WP/12/22
Surges
Atish R. Ghosh, Jun Kim, Mahvash S. Qureshi, Juan Zalduendo
© 2012 International Monetary Fund
WP/12/22
IMF Working Paper
Research Department
Surges†
Prepared by Atish R. Ghosh, Jun Kim, Mahvash Saeed Qureshi, Juan Zalduendo
January 2012
Abstract
This paper examines why surges in capital flows to emerging market economies (EMEs)
occur, and what determines the allocation of capital across countries during such surge
episodes. We use two different methodologies to identify surges in EMEs over 19802009,
differentiating between those mainly caused by changes in the country’s external liabilities
(reflecting the investment decisions of foreigners), and those caused by changes in its assets
(reflecting the decisions of residents). Global factors—including US interest rates and risk
aversion—are key to determining whether a surge will occur, but domestic factors such as
the country’s external financing needs (as implied by an intertemporal optimizing model of
the current account) and structural characteristics also matter, which explains why not all
EMEs experience surges. Conditional on a surge occurring, moreover, the magnitude of the
capital inflow depends largely on domestic factors including the country’s external financing
needs, and the exchange rate regime. Finally, while similar factors explain asset- and
liability-driven surges, the latter are more sensitive to global factors and contagion.
JEL Classification Numbers: F21, F32
Keywords: surges, capital flows, emerging market economies
Author’s E-Mail Address: aghosh@imf.org; jkim2@imf.org; mqureshi@imf.org;
jzalduendo@worldbank.org
This Working Paper should not be reported as representing the views of the IMF.
The views expressed in this Working Paper are those of the author(s) and do not necessarily
represent those of the IMF or IMF policy. Working Papers describe research in progress by the
author(s) and are published to elicit comments and to further debate.
†
We are grateful to Jonathan Ostry, and participants at the Canadian Economic Association Conference 2011; IEA
World Congress; Singapore Economic Review Conference 2011; and World Bank-Bank of Indonesia Seminar
2011, for helpful comments and suggestions, and to Hyeon Ji Lee for excellent research assistance. We are also
grateful to Camelia Minoiu for sharing the data on banking connectivity with us. Any errors are our responsibility.
2
Contents
I. Introduction ................................................................................................................................ 3 II. Empirical Strategy ..................................................................................................................... 6 III. Identifying Surges .................................................................................................................. 11 A. Methodology ....................................................................................................................... 11 B. Key Features ........................................................................................................................ 12 IV. Estimation Results ................................................................................................................. 14 A. Occurrence of Surges .......................................................................................................... 14 B. Magnitude of Flows in Surges ............................................................................................ 15 C. Asset- vs. Liability-Driven Surges ...................................................................................... 16 D. Sensitivity Analysis............................................................................................................. 18 V. What’s Driving the Surges? Some Further Exploration ......................................................... 21 VI. Conclusion ............................................................................................................................. 23 References .................................................................................................................................... 25 Appendix A: The Intertemporal Optimizing Model of the Current Account .............................. 38 Appendix B: Data and Summary Statistics .................................................................................. 39 Tables
1. Summary Statistics of Selected Variables ............................................................................... 31 2. Likelihood of Surge, 19802009 ............................................................................................. 32 3. Magnitude of Surge, 19802009 ............................................................................................. 33 4. Likelihood of Surge: by Surge Type, 19802009 ................................................................... 34 5. Magnitude of Surge: by Surge Type, 19802009 .................................................................... 35 6. Likelihood of Surge: Sensitivity Analysis ............................................................................... 36 7. Magnitude of Surge: Sensitivity Analysis ............................................................................... 37 B1.Variable Definitions and Data Sources .................................................................................. 40 B2. List of Surge Episodes with the Threshold Approach, 19802009 ...................................... 41 B3. List of Surge Episodes with the Cluster Approach, 19802009 ........................................... 42 Figures
1. Net Capital Flows to EMEs, 19802009 ................................................................................. 27 2. Surges of Net Capital Flows to EMEs, 19802009 ................................................................. 27 3. Types of Surges, 19802009 ................................................................................................... 27 4. Predicted Probabilities of Surge Occurrence ........................................................................... 28 5. Predicted Probabilities of Asset- and Liability-Driven Surge Occurrence .............................. 29
6. Binary Recursive Trees for Surge Occurrence ........................................................................ 30 B1.Colombia: Net Capital Flows to GDP, 19802009................................................................ 39 B2. Surges by Region, 19802009 .............................................................................................. 39 I. INTRODUCTION
After collapsing during the 2008 global financial crisis, capital flows to emerging market
economies (EMEs) surged in late 2009 and 2010, raising both macroeconomic challenges
and financial-stability concerns. By the second half of 2011, however, amidst a worsening
global economic outlook, capital flows receded rapidly, eliminating much of the cumulated
currency gains, and leaving EMEs grappling with sharply depreciating currencies in their
wake.1 While such volatility is nothing new—historically, flows have been episodic (Figure
1)—it has reignited questions on the nature of capital flows to EMEs. What causes these
sudden surges? What determines the allocation of flows across EMEs? And do foreign and
domestic investors behave differently when making cross-border investment decisions? In
this paper, we take up these issues; a companion paper (Ghosh et al., 2012) looks at why,
when, and how capital flow surge episodes end.
The literature on this subject has a long tradition of trying to identify global “push” and
domestic “pull” factors in determining flows to recipient economies.2 Yet, in equilibrium,
capital flows must reflect the confluence of supply and demand, so there must be both push
(supply-side) and pull (demand-side) factors, and it is hard to attribute the observed flows to
one side or the other. More meaningful, therefore, may be to consider the determinants of
changes in capital flows, which might be associated with changes in supply factors (and
declining costs of funds), or changes in demand factors (and rising costs of funds), or both
(with roughly constant costs). Moreover, from a policy perspective, large changes in capital
flows—surges—are of particular interest both because of their greater impact on the
exchange rate and competitiveness, and because they are more likely to overwhelm the
domestic regulatory framework, raising financial-stability risks. In this paper, we thus focus
on surges, and examine what factors determine their occurrence as well as magnitude.3
While it is common to think of net inflows being the result of foreigners pouring money into
the country (thereby increasing residents’ foreign liabilities), they could equally result from
the asset side—residents selling their assets abroad or simply not purchasing as many foreign
assets as before. Recent literature (Milesi-Ferretti and Tille, 2011; Forbes and Warnock,
2011) stresses the need to distinguish between these cases to better understand cross-border
capital movements—especially in advanced economies, where gross flows of assets and
liabilities dominate the net movements. Though less true of emerging markets (where net
capital flows still largely reflect changes in external liabilities), the distinction may
1
For example, in the two months following the U.S. sovereign debt rating downgrade in early August 2011, the
currencies of Brazil, Korea, Mexico, and Russia depreciated by about 10-16 percent in nominal terms, which
largely offset earlier gains that had cumulated over end-2009 and mid-2011.
2
See, for example, Chuhan et al. (1993), Taylor and Sarno (1997), Hernandez et al. (2001), and IMF (2011).
3
Fernandez-Arias (1996) and Taylor and Sarno (1997) examine the main drivers of the early-1990 surge in
capital flows to Asia and Latin America by analyzing the change in net capital flows. More recently, Reinhart
and Reinhart (2008) and Cardarelli et al. (2009) catalog capital inflow “bonanzas” in advanced and emerging
economies by focusing on large net flow observations, but mostly identify key stylized facts associated with
these episodes. Forbes and Warnock (2011) focus on large gross capital flows to differentiate between episodes
of surges, stops, flights, and retrenchment, and investigate more formally the causes of these episodes.
4
nevertheless be worth making, as liability-driven inflow surges might have different
properties from asset-driven surges, and thus call for different policy responses. For example,
it seems plausible that domestic investors would be more responsive to changes in local
conditions because of informational advantages, while foreign investors may be more
sensitive to global conditions. If so, and associating asset-driven surges with the investment
decisions of domestic residents, and liability-driven surges with those of foreigners, there
would be corresponding implications for the different types of surges.
In this paper, therefore, we also differentiate between asset- and liability-driven surges, and
compare their determinants. We do so by first identifying surges in net capital flows, and
then classifying the surge according to whether it corresponds mainly to changes in the
country’s foreign asset or liability position.4 In addition, while earlier studies have often
focused on a selected set of push and pull factors—typically ignoring the real domestic
interest rate and/or the country’s external financing needs—we systematically account for the
plausible drivers of surges, including the return differential (adjusted for expected changes in
the exchange rate), measures of risk in global markets, as well as the macroeconomic and
structural characteristics, and the external financing needs of the recipient country. A key
innovation of our study in this regard is to use an intertemporal optimizing model of the
current account to proxy for the country’s external financing needs.
We begin our empirical analysis by developing simple algorithms to identify surge episodes
in 56 EMEs over 1980–2009. We employ two methods: a “threshold” approach—net flows
(in percent of GDP) that fall in the top 30th percentile of the country’s own, and the entire
sample’s, observations; and a “clustering” approach that avoids imposing ad hoc thresholds,
and uses statistical clustering techniques on (standardized) net flows to distinguish between
surges, normal flows, and outflows. With these two methods, we identify 290 and 338 surges
(around one-fourth of the panel), respectively, roughly synchronized in the early 1980s (prior
to the onset of the Latin American debt crisis); the early 1990s (as these countries emerged
from the debt crisis); and the mid-2000s, as capital flows to EMEs recovered from the Asian
crisis and the Russian default, and then accelerated in the run-up to the global financial crisis.
The very synchronicity of surge episodes across countries suggests that global factors might
be at play. Indeed, we find this to be the case—global factors, including US interest rates,
and global risk aversion (as captured by the volatility of the S&P 500 index)—are key
determinants of whether inflow surges to EMEs will occur. At the same time, whether a
particular EME experiences a surge also depends on its own attractiveness as an investment
destination. Fundamentals, including external financing needs implied by the consumption4
Recent work by Forbes and Warnock (2011) comes closest to our study, but differs in several respects. Forbes
and Warnock identify surges as large flows of assets or liabilities, rather than on the basis of net capital flows;
as such, many of their identified surges do not correspond to periods of exceptionally large net inflows, and
therefore do not carry the same policy implications (e.g., for currency appreciation pressures). They also only
look at the determinants of surge occurrence, not of its magnitude. Finally, Forbes and Warnock comingle
advanced and emerging market economies in their sample—not surprisingly, therefore, they tend to find that
advanced economy interest rates are unimportant (as any effect of higher advanced economy interest rates in
reducing flows to EMEs is likely offset by their positive impact on flows to advanced economies).
5
smoothing optimal current account deficit, financial openness and interconnectedness, real
economic growth, and institutional quality also help determine the likelihood that the country
experiences an inflow surge. Conditional on the surge occurring, moreover, domestic “pull”
factors, including the country’s growth rate, external financing needs, and exchange rate
regime, are important in determining its magnitude. Broadly speaking, therefore, surges in
capital flows to EMEs are driven by global push factors—but where they end up depends
equally on domestic pull factors, which explains why not all countries experience a surge
when aggregate flows toward EMEs rise sharply.
Our analysis also shows that inflow surges to EMEs are mainly liability-driven—only onethird of the net flow surges correspond to changes in residents’ foreign asset transactions.
The factors driving the two types of surges turn out to be quite similar: global factors matter
for both, with lower US interest rates and greater risk appetite encouraging both foreigners to
invest more in EMEs, and domestic residents to invest less abroad. Yet some differences are
discernible. Foreign investors are equally attuned to local conditions as domestic investors,
but tend to be more sensitive to changes in the real US interest rate and global risk, and are
also more subject to regional contagion than asset-driven surges. These conclusions are
reaffirmed from a binary recursive tree analysis, which shows that global factors,
specifically, global risk, play a key role in driving large foreign inflows to EMEs.
Our findings, which are robust to different estimation methodologies, surge identification
algorithms, and model specifications, hold important policy implications. Inasmuch as surges
reflect exogenous supply-side factors that could reverse abruptly, or are driven by contagion
rather than fundamentals, the case for imposing capital controls (provided macro policy
prerequisites have been met; Ostry et al., 2011) on inflow surges that may cause economic or
financial disruption—and for greater policy coordination between source and recipient
countries—is correspondingly stronger. If the aggregate volume of capital flows to EMEs is
largely determined by supply-side factors, but the allocation of flows across countries
depends on local factors (including capital account openness), there may also be a need for
coordination among recipient countries to ensure that they do not pursue beggar-thy-neighbor
policies in an effort to deflect unwanted surges to each other.
Our contribution to the existing literature is thus three-fold. First, we focus on surges of net
capital flows, examining both why they occur, and the magnitude of the flows conditional on
their occurrence. Second, we differentiate between asset- and liability-driven surges, and
examine whether they react differently to changes in global and local conditions. Third, we
systematically account for the plausible drivers of surges—including the return differential
(adjusted for the expected exchange rate changes), and an important new proxy of the
country’s external financing needs obtained from an intertemporal optimizing model of the
current account—and complement our regression analysis with binary recursive trees.
The rest of the paper is organized as follows. Section II outlines our empirical strategy for
investigating the determinants of surge occurrence and magnitude. Section III describes how
we identify inflow surges, and documents the key features associated with surge episodes.
Section IV presents the main empirical results and sensitivity analysis. Section V further
explores the drivers of inflow surges using binary recursive trees. Section VI concludes.
6
II. EMPIRICAL STRATEGY
Growing financial integration over the past few decades, together with the evident volatility
of capital flows, has spawned a voluminous literature on the determinants of cross-border
capital flows. While early empirical studies paid particular attention to the role of interest
rate differentials (for example, Branson, 1968; and Kouri and Porter, 1974), later studies
have characterized the determinants into “push” and “pull” factors, and focused more on
evaluating the relative importance of each (for example, Chuhan et al., 1993; FernandezArias, 1996; Fernandez-Arias and Montiel, 1996; Taylor and Sarno, 1997).5 Push factors
reflect external conditions (or supply-side factors) that induce investors to increase exposure
to EMEs—for example, lower interest rates and weak economic performance in advanced
economies, lower risk aversion, and booming commodity prices. Pull factors are recipient
country characteristics (or demand-side factors) that affect risks and returns to investors, and
depend on local macroeconomic fundamentals, official policies, and market imperfections.6
Since, in equilibrium, flows must reflect the confluence of supply and demand, it is not
surprising that most studies of the level of capital flows find that both push (supply-side) and
pull (demand-side) factors matter (see, for example, Chuhan et al., 1993; Taylor and Sarno,
1997; Griffin et al., 2004; IMF, 2011; Fratzscher, 2011).7 But those that look at the change in
capital flows present a more mixed picture. Calvo et al. (1993) and Fernandez-Arias (1996)
find a dominant role of global factors, notably US interest rates, in driving capital flows to
Latin America and Asia in the early 1990s, while, for a similar sample, Taylor and Sarno
(1997) find that US interest rates and domestic credit worthiness are equally important for
changes in equity flows, but that US interest rates are much more important in driving the
short-run dynamics of bond flows.
And what about surges? The dynamics and determinants of these (exceptionally large) capital
flows may be quite different from more normal variations, but existing empirical evidence is
scant.8 The few available studies (Reinhart and Reinhart, 2008; Cardarelli et al., 2009) simply
5
See Goldstein et al. (1991) for a review of early literature on the determinants of capital flows, and Forbes and
Warnock (2011) for a more recent review.
6
Fernandez-Arias and Montiel (1996) develop an analytical framework to incorporate the effect of domestic
and global factors on capital flows by defining domestic factors as those that operate at the project/country
level, and by imposing an arbitrage condition that equates the project’s risk-adjusted expected rate of return to
the opportunity cost of the funds (which depends on creditor country conditions). Their model thus links net
capital flows with the domestic economic environment and credit worthiness, and creditor country’s financial
conditions.
7
Chuhan et al. (1993) find that the importance of push and pull factors varies across regions and types of flows.
For example, their results show that US flows to Latin America in 1988–92 were equally sensitive to pull and
push factors, but those to Asia reacted more to pull factors; and relative to equities, bond flows are more
responsive to domestic factors such as a country’s credit rating. Fernandez-Arias (1996), however, argues that
since domestic creditworthiness is closely tied to global interest rates, it is ultimately creditor country conditions
that matter more.
8
In contrast, the factors associated with large downward swings in net capital flows (in the context of sudden
stops and current account reversals) have been well explored empirically (e.g., Milesi-Ferreti and Razin, 2000;
Calvo et al., 2004; Eichengreen and Adalet, 2005).
7
present some stylized facts about the association of net flow surges with global factors such
as US interest rates, world output growth, and commodity prices, as well as with local
characteristics, notably the current account deficit and real GDP growth. Looking at gross
flows, Forbes and Warnock (2011) find that global risk, global liquidity, and global as well
as domestic real growth matter for inflow surges, but find no role of advanced economy
interest rates.9 They show, however, that the retrenchment of residents’ assets abroad is
(positively) related with interest rates in advanced economies, and with global risk and
contagion effects (through trade and financial channels).
Building on these various strands of the literature, we model both the likelihood of inflow
surges (as defined in Section III below), and their magnitude (conditional on occurrence), as
functions of: (i) the return differential, rjtd ; (ii) global push factors, xt; (iii) domestic pull
factors, zjt; and (iv) contagion, cjt:
Pr( S jt  1)  F (rjtd 1  xt1  zjt  1  cjt1 )
(1)
K jt|s jt 1  r  2  xt 2  z jt  2  cjt  2   jt
(2)
d
jt
where Sjt is an indicator variable of whether a surge in net capital flows (to GDP) occurs in
country j in period t; K jt|s jt 1 is the magnitude of the net capital flow (to GDP) conditional on
the surge, and where F(.) is assumed to follow the standard normal cumulative distribution
function so (1) can be estimated by probit, and (2) can be estimated by Ordinary Least
Squares. To address the potential endogeneity concerns of the domestic pull factors in both
(1) and (2), we substitute contemporaneous values of these variables by their lagged values.10
Since many of the structural variables (for example, capital account openness) change only
slowly, and because we are interested in the effect of global factors that will be common
across recipient countries (for example, US interest rates), we do not include country for
annual fixed effects, but control for region-specific effects and a range of country
characteristics.11
Rate of return differential
The neoclassical theory predicts that capital should respond to interest rate differentials
between countries—with capital flowing from countries with low return (capital-abundant
advanced economies) to those with high return (capital-scarce emerging economies). The
9
Forbes and Warnock’s (2011) finding that advanced economy/US interest rates are insignificant in explaining
surges may, however, be the result of their sample, which includes both advanced and emerging market
economies. Higher advanced economy interest rates would therefore have two offsetting effects: decreasing
surges to EMEs while increasing them in advanced economies, with little or no average effect in their sample.
10
If endogeneity exists, estimates from (1) and (2) will be inconsistent. We also estimate equations (1) and (2)
using the instrumental variables (IV) probit and 2SLS estimation methods, respectively, where lagged values
are used as instruments. These estimations yield very similar results, and are available upon request.
11
Estimation results for (1) and (2) with country fixed effects are also reported in the sensitivity analysis.
8
nominal interest rate differential is given by the standard uncovered interest rate parity
condition:
i djt  i jt  (it*  (e jt 1  e jt ))
(3)
where i djt is the interest differential for country j at time t, ijt is the domestic interest rate
(money market rate or treasury bill rate, according to data availability) of the emerging
economy, it* is the advanced economy interest rate (proxied by the US 3-month treasury bill
rate), and ejt is the log nominal exchange rate (an increase in ejt represents a deprecation).
Subtracting the inflation rate from both sides of (3):
rjtd  i jt  ( p jt 1  p jt )  {it*  ( pt*1  pt* )  ( pt*1  e jt 1  p jt 1 )  ( pt*  e jt  p jt )}
(4)
rjtd  rjt  rt*  q ejt 1
(5)
or
where rjtd is the real interest rate differential; pt and pt* are the log domestic and US price
levels, respectively; rjt and rt* are the domestic and US real interest rates, respectively; and
q ejt 1 is the expected real exchange rate depreciation. We proxy for the expected real
depreciation by the log difference between the current real effective exchange rate and its
long-term trend (i.e., the implied overvaluation), q ejt 1  q j  q jt , so capital flows to EMEs
should respond positively to the differential:
rjtd  rjt  (q j  q jt )  rt*
(6)
Using (6)—that is, working in terms of the real interest rate differential—is useful because
some of the EME observations include high- or even hyperinflationary periods. In the
empirical results below, we present two estimates of (1) and (2). The first variant (the
“constrained” model) includes the real-interest rate differential as defined in (6), so that the
coefficients on the individual terms ( rjt , rt* and q ejt 1 ) are constrained to be equal. The
second variant (the “unconstrained” model) includes the terms individually so the
coefficients are unrestricted, which allows to identify whether the effect of the real interest
rate differential stems mainly from the push ( rt* ) or pull ( rjt and (q j  q ) ) factors.
Global push factors
Global push factors reflect external conditions, largely beyond the control of EMEs, which
underpin the supply of global liquidity. In addition to the real US interest rate (in the
unconstrained model), we include the volatility of the Standard & Poor (S&P) 500 index, and
world commodity prices as other global push factors.12 Higher US interest rates (proxying the
12
We use the volatility of S&P 500 index returns as an alternative to the more commonly used VIX index
because the latter is only available from 1990 onwards. As a robustness check, however, we also use the: (i)
Credit Suisse global risk appetite index (RAI), which is available from 1984 onwards, and measures excess
return per unit of risk with lower (higher) values indicating periods of financial market strain (ease), and (ii)
(continued)
9
rate of return in advanced economies) are expected to reduce capital flows to EMEs.
Likewise, greater volatility of the S&P 500 index—as a measure of global market
uncertainty—is likely to reduce the surge probability for EMEs since advanced economies
are traditionally considered to be safe havens in times of increased uncertainty. Higher
commodity prices (measured as the log difference between the actual and trend commodity
price index to capture the effect of large movements in commodity prices) are likely to be
positively correlated with inflow surges inasmuch as they indicate a boom in demand for
EME exports, and perhaps the recycling of income earned by commodity exporters.
Domestic Pull Factors
For capital to flow, there must be corresponding investment opportunities in the destination
country. Early studies of private capital flows to developing countries often included the
country’s current account deficit as a measure of its financing needs (Kouri and Porter,
1974). But with the increasing importance of private (as opposed to official) flows to EMEs,
this becomes almost tautological: abstracting from changes in reserves, the current account
deficit must be (largely) financed by private capital flows, and the observed flows must
correspond to the current account deficit.
To get around this problem, and to see whether capital flows to EMEs respond to
“fundamentals,” we turn to an intertemporal optimizing model of the current account (Ghosh,
1995). In such a model, the capital inflow corresponding to the optimal current account
deficit (CADt* ) can be shown to equal the present discounted value of expected changes in
national cash flow—or the difference between GDP (Qt), investment (It), and government
consumption (Gt):13

CADt*  
j 1
E{ (Qt  j  I t  j  Gt  j )}
(1  r ) j
(7)
According to the consumption-smoothing model (7), the country has an external financing
need (that is, optimally, a current account deficit) when output is temporarily low, and/or
government consumption and investment are temporarily high (for example, in the face of a
positive productivity shock). Permanent shocks, of course, have no impact on the
(consumption-smoothing component of the) current account as the country should adjust to
such shocks. Since surges are episodes of temporarily high capital inflows, they presumably
correspond to temporary shocks to the domestic economy. Accordingly, CADt* , as defined in
(7), should be a good proxy for the country’s external financing needs that are met by surges
in net capital flows.
Even if the country does have an external financing need, it may not be met if the capital
account is closed (indeed, the derivation of (7) assumes perfect capital mobility). To capture
VXO index—the precursor of the VIX—that is available from 1986 onwards. See Table B1 for a description of
variables and data sources.
13
See Appendix A for details.
10
this possibility, we include a measure of (de jure) capital account openness in (1) and (2),
which is taken from Chinn and Ito (2008), and is based on the IMF’s Annual Report on
Exchange Arrangements and Exchange Restrictions (AREAR). Countries that are more
financially open are in principle more likely to experience a surge of capital inflows than
relatively closed economies. Regardless of de jure openness, however, a country (sovereign)
that is in arrears or otherwise in default on its external payments is unlikely to be an
attractive destination for foreign investors and is less likely to experience an inflow surge.
We therefore also include a dummy variable (based on Reinhart and Reinhart, 2008) to
capture whether the sovereign is in a debt crisis such that it is unable to make its principal or
interest payments by the due date.
Fast growing economies are more likely to experience large capital flows, not only because
of their potentially large financing needs, but also because of their greater potential
productivity and returns, as are countries with better institutional quality (Alfaro et al., 2008).
Thus, we include real GDP growth rate as well as a measure of institutional quality among
the pull factors. We also include the de facto exchange rate regime (taken from the IMF’s
AREAER) to capture the possibility that the implicit guarantee of a fixed exchange rate may
encourage greater cross-border borrowing and lending. Countries that are better integrated
with global financial markets may be more likely to receive inflows (for example, as in
Ghosh et al., 2011; and Hale, 2011)—perhaps because of lower informational costs for
foreign investors or because of more diversified sources of external financing. Therefore, we
also include a measure of the country’s financial “connectedness” as proxied by its centrality
in the global banking network (specifically, by the proportion of advanced economies that
have banks with cross-border exposure to the recipient country; Minoiu and Rey, 2011).
Finally, in the unconstrained model, we include the domestic real interest rate (which should
be positively correlated with surges), and the estimated overvaluation of the currency (which
should be negatively correlated) as separate terms based on (6).
Contagion
Another external factor, which has gained much attention in recent years, is contagion.
Recent literature finds a strong effect of contagion, particularly in the context of economic
and financial crises/sudden stops (for example, Glick and Rose, 1999; Kaminsky et al., 2001;
Forbes and Warnock, 2011), and identifies several channels (trade, financial, geographic
location, or similar economic characteristics) through which contagion may occur.14 To
capture the impact of contagion on surge likelihood, we include in (1) a regional contagion
variable defined as the proportion of other countries in the region experiencing a net capital
flow surge (and, correspondingly, in the magnitude regression (2) we include the average net
flow (in percent of GDP) to other countries in the region experiencing a surge).15
14
See Claessens and Forbes (2001) for a discussion on contagion and the possible transmission mechanisms.
In addition, in the sensitivity analysis, we also include a measure of contagion through trade
interconnectedness (defined as in Forbes and Warnock, 2011).
15
11
III. IDENTIFYING SURGES
A. Methodology
Our starting point for the empirical analysis is to identify inflow surges. A common approach
in the literature is to use thresholds—for example, Reinhart and Reinhart (2008) select a cutoff of 20th percentile across countries of total net capital flows (in percent of GDP), and
Cardarelli et al. (2009) define a surge when net private capital flows (again in proportion to
GDP) for a country exceed its trend by one standard deviation (or falls in the top quartile of
the regional distribution). In recent work, Forbes and Warnock (2011) use quarterly data on
gross capital flows, and define a surge as an annual increase in gross inflows that is more
than one standard deviation above the (five-year rolling) average and where the increase is at
least two standard deviations above the average in at least one quarter.16
There are pros and cons to defining surges in terms of net or gross inflows. On the one hand,
some financial stability risks, such as foreign currency exposure of unhedged domestic
borrowers, may depend on the country’s gross external liabilities, and as argued above, the
dynamics of liabilities may be quite different from those of assets. On the other hand, most
macroeconomic consequences of capital flows (such as exchange rate appreciation or
macroeconomic overheating) and some financial-stability risks, will be related to net, not
gross, flows. Moreover, for EMEs, net capital flows mostly correspond to changes in
liabilities, with relatively little action on the asset side.17 Indeed, the problem with using gross
flows is that many of the identified “surges” may not constitute periods of net flows, let alone
exceptionally large net flows. In this paper, therefore, we define surges in terms of the net
flow of capital but use gross flow data to distinguish between those that correspond mainly to
changes in external liabilities and those that correspond to changes in assets.
We obtain data from the IMF’s Balance of Payment Statistics, and define net capital flows as
total net flows excluding “other investment liabilities of the general government” (which are
typically official loans) and exceptional financing items (reserve assets and use of IMF
credit), expressed in percent of GDP. We identify surges using two methods. The first
approach, which follows the existing literature, is to define a surge as any year in which net
capital flows exceed some threshold value. We set the threshold at the top 30th percentile for
the country, provided the net flow (expressed in percent of GDP) also falls in the top 30th
percentile for the entire (cross-country) sample. This ensures that observations of net flows
that are large by (country-specific) historical as well as international standards are included
as surges. Likewise, observations in the bottom 30th percentile (of the country-specific as
well as the full sample’s distribution) are coded as outflows; all other observations are coded
as “normal” flows.
16
These definitions are somewhat analogous to those adopted for identifying current account reversals and
sudden stops (see Reinhart and Reinhart (2008) for a review).
17
This is in contrast to advanced economies, where net flows typically reflect largely offsetting gross flows of
much greater magnitude. For only a few EMEs (e.g., Chile), and only in recent years, have gross flows become
important.
12
Our second approach is more novel and avoids imposing ad hoc thresholds. We apply
statistical clustering techniques (specifically, k-means clustering) to group each country’s
observations on (standardized) net flows such that the within-cluster sum of squared
differences from the mean is minimized (while the between-cluster difference in means is
maximized). As a result, each observation belongs to the cluster (or group) with the nearest
mean, and clusters comprise observations that are statistically similar. Using this technique,
we group each country’s data into three clusters that we identify with: (i) surges; (ii) normal
flows; and (iii) outflows. In both approaches, we group consecutive surge observations to
form a surge episode provided they are not interrupted by a year of normal flows or outflows.
While the particular choice of algorithm to identify surges inevitably involves trade-offs, our
approaches have the advantage of ensuring uniform treatment across countries while still
allowing significant cross-country variation in the absolute threshold of a surge.18 The use of
two, wholly independent approaches also gives confidence about the robustness of the
obtained results. As with other empirical studies, however, dating the beginning and the end
of surges is not always clear cut since the strict application of any algorithm to identify
surges runs the risk of omitting at least some observations of relatively large net capital flows
that may otherwise be part of an episode. We therefore also construct a one-year window
around the identified episodes, including the immediate pre- and post-surge years (provided
the net capital flow is positive in these years), and check the robustness of our estimation
results to these extended episodes.
B. Key Features
We apply the threshold and cluster approaches to a sample of 56 EMEs using annual data for
the period 1980–2009.19 There is considerable overlap in the resulting surge observations,
with a correlation of about 0.8 between them, although the threshold approach yields
somewhat fewer, but larger, surges.20 For example, Figure B1 shows the identified surge
observations for Colombia using the two approaches. There are 3 observations of net capital
flow to GDP for Colombia that are in the top 30th percentile of the country-specific
18
In our approaches, the country-specific cut-off for identifying surges remains constant over the sample period,
ensuring that capital inflows that are exceptionally large (in percent of GDP) are always coded as surges. This is
in contrast to methods that use deviations from rolling averages, which may take better account of drifts in the
volatility of capital flows, but may not code large capital flow observations as a surge if the large inflows have
persisted for a few years. Conversely, rolling methods may identify a capital inflow as a “surge” even though it
is small in absolute terms (and therefore of little macroeconomic consequence), if flows have been low for a
while but then there is a small jump in the series.
19
We use annual data because most of the explanatory variables are not available at quarterly frequency for
EMEs, especially for the early years of our sample. The list of countries included in the sample is motivated by
the EMEs covered in the IMF’s Early Warning Exercise (IMF, 2010). Tables B2 and B3 list the countries
included our sample, as well as the surge episodes obtained from the threshold and cluster approaches.
20
A comparison of our surge episodes with those of Reinhart and Reinhart (2008) and Cardarelli et al. (2009)
suggests broad overlap, particularly for “well-known” episodes, though there are some differences in the
duration of the episodes. For the countries included in our sample, the correlation between our threshold-based
surge observations with those in Reinhart and Reinhart, and Cardarelli et al. is 0.3 and 0.4, respectively (while
the correlation between the surge series in these two studies is around 0.3).
13
distribution, as well as in the top 30th percentile of the overall distribution of net capital flows
to GDP, and hence are coded as surges under the threshold approach. Through the cluster
analysis, however, we obtain 9 surge observations, half of which are large from the country’s
historical (but not from a global) perspective. In what follows, we focus on the results when
surges are defined using the threshold approach, which are more extreme as they reflect both
country-specificity and international uniqueness, reserving the cluster-identified surges for
our robustness checks.
Under the threshold approach, we obtain 290 surge observations (which yield 149 surge
episodes), the majority of which are in Eastern Europe and Latin America. The average
duration of each episode is around 2 years, while the average net capital flow during the
episode is about 10 percent of GDP. As a proportion of GDP, the largest surges are actually
in the Middle East and African countries (around 12 percent of GDP, perhaps because of
large resource extraction investment projects), followed by emerging Europe. Surges have
become more frequent in recent years with the share of surge observations rising from about
10 percent in the 1980s to more than 20 percent in the 1990s, and to almost 30 percent in the
last decade (Figure 2). Similar patterns are obtained when cluster analysis is applied—338
surge observations (grouped into 168 surge episodes) are identified, most of which coincide
with those from the threshold approach.
Classifying by the type of surge shows that the majority (more than two-thirds) are driven by
an increase in residents’ liabilities (liability-driven) rather than by a decline in the holdings of
their assets abroad (asset-driven).21 Asset-driven surges outnumber liability-driven surges in
only two out of the 30 years of our sample—1982 and 2008, both of which are crisis years
(Figure 3, panel a).22 Moreover, looking at surge episodes, nearly all begin with a liabilitydriven surge, suggesting that as foreign investment starts flowing into an EME, domestic
investors follow suit, repatriating foreign assets in order to invest at home. On average,
liability-driven surges are also somewhat larger than asset-driven surges, though the
difference is not statistically significant (Figure 3, panel b).
An initial snapshot of the occurrence and magnitude of inflow surges suggests three
noteworthy points. First, surges seem to be synchronized internationally (Figure B2),
generally corresponding to “well-established” periods of high global capital mobility—the
early 1980s (just before the Latin American debt crisis), the mid-1990s (before the East
Asian financial crisis and Russian default), and the mid-2000s in the run-up to the recent
financial crisis—suggesting that common factors are at play. Second, even in times of such
global surges, not all EMEs are affected. In fact, the proportion of EMEs experiencing an
inflow surge in any given year never exceeds one-half of the sample, with some countries
experiencing them repeatedly. As such, conditions in the recipient countries must also be
21
To determine whether a surge is driven by an increase in residents’ liability or asset transactions, we use data
on total liabilities (gross inflows) and total assets (gross outflows) from the Balance of Payments. Thus, when a
net capital flow surge corresponds to a larger increase in domestic residents’ liabilities vis-à-vis a reduction in
their assets, it is identified as liability-driven, while it is defined as asset-driven when the converse holds.
22
This observation is in line with the well-established drawdown on residents’ foreign assets during crises
(Milesi-Ferretti and Tille, 2011).
14
relevant. Third, there is considerable time-series and cross-sectional variation in the
magnitude of flows conditional on the occurrence of a surge. For example, Asian countries
experienced the largest surges (in proportion of GDP) during the 1990s wave of capital
flows, whereas emerging Europe experienced the largest surges in the mid-2000s wave.
Thus, both global and domestic factors appear to be relevant in determining surges—perhaps
global factors driving the overall volume of flows to EMEs, and domestic factors influencing
their allocation.
What are these factors? A simple tabulation of explanatory variables during surge, normal,
and outflow periods suggests a number of global push and domestic pull factors may be
relevant (Table 1). During surge periods, the US real interest rate and global market
uncertainty (S&P 500 index volatility) are lower, while commodity prices are higher, than at
other (normal or outflows) times. Turning to domestic factors, when experiencing surges,
recipient countries tend to have larger external financing needs, and faster output growth, as
well as more open current and capital accounts (with greater financial interconnectedness),
and stronger institutions.
IV. ESTIMATION RESULTS
The statistics reported in Table 1 are suggestive of the factors that might determine when and
whether a country experiences an inflow surge. In what follows, we examine more formally
the determinants of the occurrence and magnitude of surges. Below, we also split surges
according to whether they are asset or liability-driven and conduct various robustness checks
on our results.
A. Occurrence of Surges
We begin by estimating the “constrained” variant of the surge occurrence probit model
specified in (1), where the real interest rate differential (adjusted for the expected real
exchange rate depreciation) enters as a single composite variable (Table 2, cols. [1]-[5]).
According to the estimates, a higher real interest rate differential raises the likelihood of an
inflow surge, though the coefficient only becomes statistically significant when domestic pull
factors are taken into account. Greater global market uncertainty (volatility of the S&P 500
index) has a strong dampening effect on the probability of a surge of capital to EMEs,
presumably because—at least traditionally—these countries have not been viewed as safe
havens at times of heightened uncertainty and risk aversion. Conversely, commodity price
booms, which likely signal higher global demand for EME exports, are positively correlated
with inflow surges, as is regional contagion (though the latter becomes statistically
insignificant after controlling for domestic pull factors). Although individual coefficients are
highly statistically significant, these global factors have limited explanatory power: the
pseudo-R2 (which compares the log likelihood of the full model with that of a constant only
model) is 8 percent, and the probit sensitivity (proportion of surges correctly called) is about
14 percent.
Turning to domestic pull factors, the external financing need implied by the optimal
consumption-smoothing current account is highly significant as is real GDP growth in the
recipient country. Countries with fewer capital account restrictions, that are better connected
15
(in the sense of more sources of cross-border loans), or that have stronger institutions are also
significantly more likely to experience inflow surges. Countries with more flexible exchange
rate regimes or that are in default are less likely to experience inflow surges, though neither
variable is statistically significant. Adding these pull factors more than doubles the pseudoR2 to 20 percent and raises the sensitivity to 27 percent.
The right-hand panel of Table 2 (cols. [6]-[10]) reports the corresponding estimates when the
real interest rate differential is not constrained to enter as a single term so that the US real
interest rate, domestic real interest rate, and estimated real exchange rate overvaluation enter
separately. Doing so shows that much of the effect of the real interest rate differential is
through the US real interest rate: evaluated at the mean of other explanatory variables, a 100
basis point rise in US real interest rates would lower the likelihood of an inflow surge by 3
percentage points (where the unconditional probability of a surge in the estimated sample is
22 percent). Real exchange rate overvaluation lowers the estimated likelihood of a surge,
though the coefficient is not statistically significant when the full set of domestic factors is
added (Table 2, cols. [9]-[10]), while the domestic real interest rate, though positive, is not
statistically significant in any of the specifications. Most of the other variables are of similar
magnitude and statistical significance to those estimated under the restricted variant. Overall,
the model correctly calls some 80 percent of the observations, and almost 30 percent of the
surge observations, with a pseudo-R2 of 21 percent.
To put the estimated effects in perspective, Figure 4 plots the implied probability of a surge
evaluated around the means of the explanatory variables based on the estimates reported in
column (10). Against an unconditional probability of 22 percent, a one standard deviation
shock to the volatility of the S&P 500 index lowers the predicted surge probability by about 3
percentage points, while the corresponding shock to the commodity price index raises the
surge probability by about 7 percentage points. Turning to domestic macroeconomic factors,
a one percentage point increase in the country’s real GDP growth rate, or a one percent of
GDP increase in its external financing needs, raises the predicted likelihood of a surge by
about 1 and 3 percentage points, respectively. On capital account openness and the
institutional quality index, moving from the sample median to the 75th percentile raises the
predicted probability of a surge by some 4 to 5 percentage points, respectively.
B. Magnitude of Flows in Surges
The probit estimates above give the likelihood of experiencing an inflow surge, but the
magnitude of the capital flow during a surge also varies considerably (ranging from 4 percent
of GDP to about 54 percent of GDP as shown in Table 1). Is it possible to say anything about
the size of the surge conditional on its occurrence? Table 3 reports the estimation results for
the surge magnitude regression (2), where the dependent variable is net capital flow
(expressed as a proportion of GDP), and the sample comprises only the surge observations.
Again, we present the constrained model in which the real interest rate differential enters the
regression as a single term, and the unconstrained model where the US real interest rate,
domestic real interest rate, and overvaluation are allowed their own coefficients. The real
interest rate differential is statistically insignificant, while the global factors appear to play a
more limited role. A 100 basis point decline in the real US interest rate is associated with
16
almost 1 percent of GDP larger capital flows (Table 3, cols. [7]-[12]), but commodity price
booms and S&P 500 index volatility have mostly insignificant effects, suggesting that these
factors act largely as “gatekeepers”—capital surges toward EMEs only when these global
conditions permit, but once this hurdle is passed, the volume of capital that flows is largely
independent of it. An interesting finding is that of a negative (and in the unconstrained
model, statistically significant) effect of the regional contagion variable, which most likely
indicates that an increase in the average flow to other countries in the region implies less
capital left to be allocated to the country in question.
Since countries that experience a surge already share the macroeconomic and structural
characteristics identified above, several of the domestic pull factors are statistically
insignificant. Nevertheless, the nominal exchange rate regime, real exchange rate
overvaluation, and external financing needs of the country are all highly statistically
significant. A one-percent of GDP increase in the estimated external financing need is
associated with one-third of one percent of GDP higher capital inflows, while 10 percent
overvaluation of the real exchange rate is associated with about 2 percent of GDP lower net
capital flows. Other factors equal, a country with a pegged exchange rate would experience 3
percent of GDP larger capital flows during a surge than if it had a more flexible exchange
rate regime. Finally, countries with more open capital accounts appear to experience larger
surges: moving from the 25th percentile of the sample’s capital account openness index to the
75th percentile is associated with 1 percent of GDP higher capital inflows during a surge.
Overall, these findings are consistent with, but go beyond, the results of previous studies, and
help to explain the stylized facts noted in Section III. Specifically, the finding that the
likelihood of surge occurrence is influenced strongly by global factors—notably, the US
interest rates, as argued by Calvo, Liederman and Reinhart (1993) and Reinhart and Reinhart
(2008), and global risk—explains the synchronicity of surges across regions, and highlights
that sudden changes in these factors could trigger large swings in capital flows. Certain
macroeconomic (in particular, growth performance and the external financing need), and
structural characteristics (notably, financial openness and institutional quality), are also
important for a surge to occur, which explains why not all countries experience a surge when
in aggregate capital is flowing toward EMEs. Further, among the countries that experience a
surge, the magnitude of the flow appears to be driven not only by the real US interest rate
and external financing need, but also by the exchange rate regime and financial openness,
with countries that have less flexible exchange rate regimes or those that are more financially
open experiencing larger surges.
C. Asset- vs. Liability-Driven Surges
Does the nature of the surge matter? In other words, are the global and domestic factors
identified above equally important for surges that mainly reflect changes in residents’ assets
(asset-driven surges) and those caused by changes in their liabilities (liability-driven surges)?
To examine this question, we re-estimate (1) and (2), but define the surge as being either
17
asset or liability-driven.23 The top panel in Tables 4 and 5 presents the results for the
constrained model for the two types of surges, while the bottom panel presents estimates for
the unrestricted model with the real US and domestic interest rates and real exchange rate
overvaluation included as separate terms.24
The results for the probit model show that the real interest rate differential raises the
likelihood of both asset and liability-driven surges, but the impact is statistically significant
for the latter only (Table 4, columns [5]-[10]). Increased global market uncertainty however
matters strongly for both types of surges, such that in times of increased global market
uncertainty, foreign as well as domestic investors exit EMEs and prefer to invest in safe
haven countries.25 Nevertheless, foreign investors appear to be more sensitive to global
market uncertainty: a one standard deviation shock to the S&P 500 index volatility reduces
the likelihood of a liability-driven surge by 4 percentage points compared to 1 percentage
point for asset-driven surges. Asset-driven surges are more likely when commodity prices are
booming—whereas commodity prices have no discernible impact on liability-driven surges.
By contrast, liability-driven surges are more subject to regional contagion.
Among the domestic pull factors, the external financing need, real economic growth, capital
account openness, and institutional quality matter for both types of surges. Liability-driven
surges, however, are more sensitive to the recipient country’s external financing need and
economic growth prospects—such that a one percentage point increase (at mean values),
raises the likelihood of a liability-driven surge by about an additional 1 and 0.5 percentage
point, respectively, as compared to an asset-driven surge (Figure 5). The impact of capital
account openness is somewhat similar on both types of surges—moving from the 25th
percentile of the capital account openness index to the 75th percentile raises asset- and
liability-driven surge probabilities by about 3 percentage points. The strong impact of capital
account openness on asset-driven surges is intuitive as only when capital flows are
liberalized (and can leave the national jurisdiction) in the first place, could they be retrenched
from abroad and invested in the domestic economy. Financial interconnectedness, however,
has a more pronounced impact on liability-driven surges—indicating that EMEs with greater
financial linkages are more likely to experience large foreign capital flows.
The results of the unconstrained model show that real US interest rates matter significantly
for the occurrence of both asset and liability-driven surges, though the impact is larger for the
latter—a 100 basis points increase in the real US interest rate (evaluated at mean values)
lowers the predicted probability of a liability-driven surge by about 2 percentage points, and
23
In these estimations, the comparison of each surge type is with the nonsurge observations; hence, the
observations for the other type of surge are excluded from the sample. As mentioned in Section III, asset-driven
surges constitute only 7 percent of the full sample (about one-third of the surge observations); thus, results
pertaining to these estimations may be interpreted with caution.
24
The sample size for both asset and liability-driven surges is different as in each case we exclude the
observations for the other type of surge from the estimation sample.
25
Forbes and Warnock (2011) find that an increase in global risk raises the likelihood of retrenchment (that is, a
decline in the residents’ assets held abroad). Considering that their sample predominantly comprises advanced
economies, their finding does not appear to be at odds with ours.
18
that of an asset-driven surge by 1 percentage point. Interestingly, asset-driven surges appear
to react more strongly to changes in the real domestic interest rate, while liability-driven
surges respond more to expected changes in the exchange rate with greater real exchange rate
overvaluation (and hence expected depreciation) making liability-driven surges less likely.
For the other factors, the signs and magnitude of the estimated effects from the unconstrained
model are very similar to those obtained above from the constrained model.
In terms of the magnitude of flows during surges, as before, several domestic
macroeconomic and structural characteristics are statistically insignificant because by
definition countries are sufficiently similar to have experienced a surge. Nevertheless, the
results from the constrained model show that the real interest differential and other global
factors are statistically insignificant for bother asset- and liability-driven surges. Domestic
factors, however, do matter. The size of inflows received is larger if nominal exchange rate
regimes are less flexible and capital accounts are more open (Table 5, panel a). Thus, a
country with a pegged exchange rate experiences about 2 and 5 percent of GDP larger capital
flows during asset- and liability driven surges, respectively, than if it had a more flexible
exchange rate regime. Likewise, moving from the 25th percentile of the capital account
openness index to the 75th percentile raises the size of the surge by about 1-2 percent of GDP
for asset- and liability-driven surges. The external financing need and regional contagion,
however, strongly impact the magnitude of liability-driven surges only such that larger
external financing needs, and smaller inflows to other countries in the region imply larger
surges. Moreover, the results from the unconstrained model (Table 5, panel b) show that real
US interest rates, and real exchange rate overvaluation also strongly affect the magnitude of
liability-driven surges—specifically, a 100 basis point increase in the real US interest rate,
and a 10 percentage point real exchange rate overvaluation imply lower inflows by about 1
and 2 percent of GDP, respectively.
The results for both the occurrence and magnitude of surges suggest that while asset- and
liability-driven surges have many common factors, there are also some important differences.
In particular, liability-driven surges are more sensitive to global factors and to contagion, but
are also more responsive to the external financing needs of the country and dependent on its
financial interconnectedness. Inasmuch as liability-driven surges reflect the investment
decisions of foreigners, these findings make intuitive sense.
D. Sensitivity Analysis
To check the robustness of our estimates reported above, we conduct a range of sensitivity
tests below, which pertain to the dating and coverage of surge episodes, our alternative
methodology for identifying surges (cluster analysis), model specification (alternative
proxies and additional regressors), and sample period.
Extended episodes
Pinning down the exact timing (beginning and end) of surge episodes is not always
straightforward. Thus, while our surge episodes largely overlap (for at least one year) with
episodes identified in other studies (e.g., Reinhart and Reinhart, 2008; and Cardarelli et al.,
19
2009), they do not coincide completely (nor do surge episodes identified in other studies
correspond exactly with each other). In general, strict application of any algorithm to identify
surges runs the risk of omitting at least some observations of relatively large net capital flows
that in reality are probably part of the same episode but that do not quite meet the criteria.
To address these concerns, we construct a one-year window around the identified episode,
including the year immediately before and immediately following the surge episode
(provided the net private capital flow in those years is positive), and re-estimate all
specifications using the extended surge variable.26 Tables 6 and 7 (col. [1]) present the
estimation results for this exercise, which largely support the findings reported in Tables 2
and 3, respectively. Specifically, for surge likelihood, the US real interest rate, global market
uncertainty, and commodity prices are all significant—while domestic factors such as the
external financing need, real GDP growth rate, capital account openness, financial
interconnectedness, and institutional quality are also strongly significantly. The main
difference with the previous results is that the domestic real interest rate and the sovereign
default dummy now become statistically significant. For surge magnitude, as before, the
external financing need, a less flexible exchange rate regime, and expected real exchange rate
change matter, as do the real US interest rate, and the amount of inflow received by other
countries experiencing a surge in the region. There is also some evidence for extended
episodes that commodity price booms lead to larger surges.
Cluster analysis
The estimation results for surges obtained through cluster analysis—which identifies about 4
percent more surge observations in the sample—also present a somewhat similar picture to
those obtained from the threshold approach (Tables 6 and 7, col. 2). The impact of both
global and domestic factors is generally comparable to earlier estimates in terms of statistical
significance and magnitude—with the only notable exception being the capital account
openness variable, which loses its statistical significance in Table 6. The weakening of the
estimated impact of capital account openness however may reflect the fact that it matters
more pronouncedly for extreme net flows as identified by the threshold approach. The surge
magnitude regressions for the cluster approach also present a similar picture to that obtained
above, and show that among the domestic pull factors, the external financing need, exchange
rate overvaluation, and exchange rate regime matter, along with the real US interest rate.
Alternative specifications
While the estimations reported in Tables 2 and 3 include a range of global and domestic
factors, other proxies and additional variables could also be employed. To check the
sensitivity of our results to alternative variable definitions of global factors, we use the 10year US government bond yield (instead of the 3-month US Treasury bill rate); and replace
our S&P 500 volatility index with the Credit Suisse risk appetite index (higher values
26
The number of extended surge episodes is smaller than before (106) as several of the one or two-year apart
surge episodes combine to form one long episode. The total number of surge observations is however larger
(496).
20
indicating periods of financial ease, and greater investor risk appetite), and the VXO index
(higher values indicating higher volatility in international markets, and lower risk appetite).
The revised estimation results reported in Table 6 (cols. [3]-[5]) show that using these
alternate proxies does not have much impact on the results—both lower US interest rates and
greater risk appetite (as measured by the Credit Suisse index) raise the likelihood of surge
occurrence. The estimated coefficient of the VXO index is however negative but statistically
insignificant, while the results of all other variables remain the same as before.
Columns [6]-[11] (in Tables 6 and 7) include additional pull variables in our general
specification to capture the effect of other potentially important domestic characteristics,
while column [12] includes country fixed effects. For example, a country’s trade openness,
and financial sector development may increase its attractiveness as an investment destination,
and boost the likelihood and magnitude of surges. The results show that indeed trade
openness significantly raises the likelihood of surge occurrence, but the proxies for financial
sector development and soundness (such as stock market capitalization, private sector credit
to GDP, and banks’ return on equity) are statistically insignificant. We also do not find a
strong impact of contagion through trade relationships on surge likelihood/magnitude. The
inclusion of these variables however does not affect much the estimated magnitude and
significance of the other pull factors in the regressions.27
Although, as shown in Figure 1, a surge in international capital flows to the EMEs occurred
in the early 1980s (as a continuation of the surge in late 1970s), surges in later years—
particularly post-Latin American debt crisis—have been larger (in both absolute and relative
to GDP terms), and have also involved more countries (Figure 4, panel b).28 To examine if
the role of global and local factors in later years has been somewhat different, we re-estimate
our regressions for the 1990–2009 sample. The results summarized in column [13] (of Tables
6 and 7) show a largely unchanged impact of global and local factors, with the exception of
the capital account openness index, which is statistically insignificant for this period.
Finally, the surge occurrence probit has a large number of zero observations in the dependent
variable (about 80 percent of the observations in the estimated sample are zero). By
construction, the probit model specifies that the distribution of F(.) in equation (1) is normal,
and symmetric around zero. If however the distribution of the dependent variable is skewed,
applying the complementary log-log model—which is asymmetric around zero—may be
preferable.29 Table 6 (last column) presents the estimation results for the most general
specification with the complementary log-log method (with clustered standard errors at the
country level). The obtained results however remain very similar to those reported above.
27
We also include proxies for fiscal balance to GDP, public debt to GDP, and the political regime in place, but
find these variables to be statistically insignificant, while their inclusion does not alter the other results. All
results are available upon request.
28
Several studies (e.g., Chuhan et al., 1993; and Taylor and Sarno 1997) note that the composition of flows in
the surge of 1990s and later years has also been different with a pronounced increased in portfolio flows.
29
Specifically, the complementary log-log model specifies that
zit′γ µi .
x ′β
z ′γ
µ
1
exp exp x ′ β
21
V. WHAT’S DRIVING THE SURGES? SOME FURTHER EXPLORATION
While the empirical analysis conducted thus far identifies the key factors that contribute to
net capital flow surges in EMEs, it also highlights the possibility that both global and
domestic factors may somewhat interact with each other to produce the cross-sectional and
time-series pattern of surges that we observe in the data. A country may thus be more
susceptible to receive capital inflows because of strong structural characteristics or large
external financing needs, but only experience a surge when global conditions permit. In
principle, the probit model estimated above—which gives the marginal effect on the
probability of a surge for each of the explanatory variables (holding the other variables
constant at their mean values)—could be modified to include such interactions; but in
practice, this becomes extremely difficult when several explanatory variables are being
considered simultaneously and possible threshold effects are unknown.
We therefore complement our probit analysis with a decision-theoretic classification
technique—known as binary recursive tree—that readily allows for interactions between the
various explanatory variables, fleshing out any context-dependence and threshold effects in
the data. Formally, a binary recursive tree is a sequence of rules for predicting a binary
variable on the basis of a vector of explanatory variables such that at each level, the sample is
split into two sub-branches according to some threshold value of one of the explanatory
variables. The explanatory variable (and the corresponding threshold value) used for splitting
is the one that best discriminates between the groups that constitute the binary variable, as
specified a priori by some criterion.30 The splitting is repeated along the various sub-branches
until the specific criterion is not met, and a terminal node is reached. The technique thus
establishes orderings among explanatory variables such that a variable that appears toward
the top of the tree could be considered as somewhat more important in distinguishing
between the surge and no-surge cases.
Based on the results of the probit model, we include the global factors as well as the most
statistically dominant domestic (macroeconomic and structural) factors to construct the
binary tree (specifically, the optimal current account balance (to GDP), real GDP growth
rate, capital account openness, and institutional quality). Figure 6 (panel a) presents the
binary recursive tree obtained for the full sample, where the dependent variable is the
occurrence of a surge. The tree turns out to be rather simple, and shows that large external
financing need is often the proximate cause for receiving large net flows—countries with
optimal current account deficits larger than 1.3 percent of GDP are more than twice as likely
to experience a surge as countries with smaller deficits. However, global risk features
notably, and countries with large external financing needs are 2½ times more likely to
experience a surge when global market volatility (or risk aversion) is low. The tree correctly
classifies about 81 percent of the sample; 16 percent of the surge observations and 98 percent
30
While several algorithms are available to search for the best split, we employ the Chi-squared Automatic
Interaction Detector (CHAID) that relies on the chi-squared test to determine the best splits (see Kass (1980) for
details). Implementation of CHAID is undertaken using the SIPINA classification tree software.
22
of the no-surge observations.31 Note that nothing prevents the algorithm from further splitting
the tree (using any of the explanatory variables); however, given the stopping rule for the
algorithm, the improvement in the fit is not sufficient to justify the additional complexity of
the tree.
Next, we construct separate binary trees for asset- and liability-driven surges. The resulting
trees, with the conditional probabilities of a surge at each node, are illustrated in Figure 6
(panel b). The first variable used for splitting the sample for asset-driven surges turns out to
be the optimal current account balance (to GDP) at the threshold value of about 1.3 percent
of GDP. The conditional probability of an asset-driven surge in countries with external
financing needs larger than this threshold value (the left hand branch of the tree) is 17
percent, versus 5 percent for countries that have relatively smaller financing needs.
Continuing along the left hand branch of the tree, the second node again depends on the
optimal current account balance such that countries (in our sample) with external financing
needs in excess of 9 percent of GDP are four times more likely to have an asset-driven surge
than otherwise. The next important variable along the same branch is real GDP growth rate,
with a threshold value of 6 percent, and a conditional probability of surge of about 86 percent
for countries that exceed the threshold (versus 20 percent for countries below the threshold).
Moving on to the right hand branch of the tree (that is, countries with optimal current account
deficits smaller than 1.3 percent of GDP), we see that it is institutional quality that matters—
countries with smaller financing needs, but in the top 30th percentile of the institutional
quality index have a much higher likelihood of an asset-driven surge.
While it is mainly the domestic factors that seem to be the key drivers of asset-driven surges,
global factors—specifically, global risk—dominate in explaining the occurrence of liabilitydriven surges. Thus, when global market volatility is low (in the bottom 20th percentile of the
S&P 500 index volatility), the conditional probability of a surge through an increase in
residents’ liabilities is 31 percent (versus 12 percent when global market volatility/risk
aversion is higher). However, once low global market volatility is taken into account (the left
hand side of the tree), countries with large external financing needs are almost thrice as likely
to experience a surge (conditional surge probability is 66 percent with optimal current
account deficit larger than 1.4 percent of GDP, versus 22 percent otherwise). Along the right
hand side of the tree (when global market volatility is high), again the external financing
need is what matters—countries with optimal current account deficits in excess of about 0.2
percent of GDP are thrice as likely to experience a surge than those with smaller deficits.
However, among the latter, countries with a stronger institutional quality are much more
likely to attract foreign inflows than otherwise.
Consistent with the probit analysis, our binary tree algorithms thus indicate that global and
domestic factors matter for surge occurrence. But, in addition, they show that there may be
31
Since the binary tree is generated here primarily for descriptive purposes, we use the full sample to determine
the splitting criteria. An alternative way, preferable for predictive analytics, is to “cross-validate” by splitting
the sample randomly into a core sample used for generating the tree, and a smaller test sample that is used for
out-of-sample forecasts. To test the robustness of our results, we split the sample into 80 (core sample), and 20
percent (test sample); nevertheless, a similar tree is generated.
23
particular interactions between these factors, which could differ based on the surge type, such
that liability-driven inflows are much more likely to be triggered by global factors, notably,
global uncertainty and market conditions; but asset-driven surges respond more to local
conditions.
VI. CONCLUSION
This paper examines the drivers of large net capital flows—or surges—to EMEs by explicitly
distinguishing between their occurrence and magnitude, and by differentiating between assetand liability-driven surges, which likely correspond to the investment decisions of residents
and foreigners, respectively. We use simple algorithms—based on thresholds and cluster
analysis—to quantitatively identify such episodes, which while certainly not exhaustive,
capture most known episodes, and allow us to distinguish between surges and more normal
periods of net capital flows.
Our descriptive analysis based on a sample of 56 EMEs over 1980–2009 indicates that surges
are synchronized internationally, and have become more frequent in recent years.
Nevertheless, even when surges occur globally, they are relatively concentrated—with never
more than half of the EMEs in the sample experiencing them at one point of time, and some
countries experiencing them repeatedly. The amount of capital received in a surge varies
considerably across countries, while most (over two-thirds) of the surges to EMEs are driven
by an increase in residents’ liabilities rather than by a decline in their foreign assets.
The picture that emerges from our regression analysis is one in which global push factors,
notably, the real US interest rate and global market uncertainty, determine whether there will
be a surge of capital flows towards EMEs generally, which helps to explain why surges are
synchronized internationally, and why they recur. Of course, a country that has no need for
capital or that is an unattractive destination for investors will not receive inflows even if there
is a global surge of capital to EMEs; hence pull factors, particularly, economic performance,
the external financing need, capital account openness, and institutional quality, do play some
role in the occurrence of a surge, and explain why some countries do (and others do not)
experience surges. Conditional on the surge occurring, moreover, domestic pull factors,
including the exchange rate regime, are important in determining its magnitude.
Our results also indicate that domestic and foreign investors respond to both global and local
factors such that lower US interest rates encourage capital to flow to EMEs while increased
global market uncertainty leads capital to flow out towards traditional safe-haven assets.
Foreign investors, however, appear to be more sensitive to global conditions than domestic
investors, with a change in the real US interest rate and global uncertainty raising the
predicted likelihood of liability-driven surges somewhat more than for asset-driven surges.
These results are reaffirmed by a binary recursive tree analysis which shows that liabilitydriven inflows are much more likely to be triggered by global factors, notably, global
uncertainty and market conditions; but asset-driven surges respond more to local conditions.
Overall, our findings provide a better understanding of large upward swings in capital flows
to EMEs, and suggest that inasmuch as they reflect exogenous supply-side factors, that could
24
reverse abruptly, the case of EMEs for imposing capital controls in the face of inflow surges
that may cause economic or financial disruption may be correspondingly stronger (provided
macro policy prerequisites have been met; Ostry et al., 2011). Conversely, however, if the
aggregate volume of capital to EMEs is driven by supply-side factors, but local factors
(including capital account openness) also play a role in determining the allocation, then there
may be a need for greater coordination between both source and recipient countries (to
ensure that there are no unintended consequences of the policy actions of the former); and
among EMEs to ensure that they do not pursue beggar-thy-neighbor policies against each
other. Further, while the drivers of asset and liability-driven surges may be largely similar,
policy responses may need to be adjusted to the type of surge—for example, prudential
measures might be more important for dealing with financial-stability risks caused by assetdriven surges, but capital controls on inflows may be an additional option for liability-driven
surges.
25
REFERENCES
Alfaro, L., S. Kalemli-Ozcan, V. Volosovych, 2008, “Why Doesn’t Capital Flow from Rich
to Poor Countries? An Empirical Investigation,” Review of Economics and Statistics, Vol. 90
(2), pp. 347-368.
Branson, W., 1968, Financial Capital Flows in the U.S. Balance of Payments (Amsterdam:
North-Holland).
Calvo, G., A. Izquierdo, and L. Mejia, 2004, “On the Empirics of Sudden Stops: The
Relevance of Balance Sheet Effects,” NBER Working Paper 10520, (Cambridge: NBER).
Calvo, G., L. Leiderman, and C. Reinhart, 1993, “Capital Inflows and Real Exchange Rate
Appreciation in Latin America: The Role of External Factors,” IMF Staff Papers, Vol. 40 (1),
pp. 108-151.
Cardarelli, R., S. Elekdag, and A. Kose, 2009, “Capital Inflows: Macroeconomic
Implications and Policy Responses,” IMF Working Paper 09/40 (Washington DC: IMF).
Chinn, M., and H. Ito, 2008, “A New Measure of Financial Openness,” Journal of
Comparative Policy Analysis, Vol. 10(3), pp. 309-322.
Chuhan, P., S. Claessens, and N. Mamingi, 1993, “Equity and Bond Flows to Latin America
and Asia: The Role of Global and Cuntry Factors,” World Bank Policy Research Working
Paper 1160 (Washington DC: World Bank).
Claessens, S., and K. Forbes, International Financial Contagion (Boston: Springer).
Eichengreen, B., and M. Adalet, 2005, “Current Account Reversals: Always a Problem?”
NBER Working Paper 11634 (Cambridge, MA: National Bureau of Economic Research).
Fernandez-Arias, E., 1996, “The New Wave of Private Capital Inflows: Push or Pull?”
Journal of Development Economics, Vol. 38(2), pp. 389-418.
Fernandez-Arias, E., and P. Montiel, 1996, “The Surge in Capital Inflows to Developing
Countries: An Analytical Overview,” World Bank Economic Review, Vol. 10 (1), pp. 51-77.
Forbes, K., and F. Warnock, 2011, “Capital Flow Waves: Surges, Stops, Flight and
Retrenchment,” NBER Working Paper 17351 (Cambridge: NBER).
Fratzscher, M., 2011, “Capital Flows, Push Versus Pull Factors and the Global Financial
Crisis,” NBER Working Paper 17357 (Cambridge: NBER).
Ghosh, A., 1995, “International Capital Mobility amongst the Major Industrialised Countries:
Too Little or Too Much?” Economic Journal, Vol. 105(428), pp. 107-128.
Ghosh, A., M. S. Qureshi, and J. Zalduendo, 2012, “Crashes,” IMF Working Paper,
forthcoming.
Ghosh, S., N. Sugawara, and J. Zalduendo, 2011, “Banking Flows and Financial Crisis—
Financial Interconnectedness and Basel III Effects,” World Bank Policy Research Working
Paper 5769 (Washington DC: World Bank).
26
Goldstein, M., D. Mathieson, and T. Lane, 1991, “Determinants and Systemic Consequences
of International Capital Flows,” IMF Occasional Paper 77 (Washington, DC: IMF).
Griffin, J., F. Nardari, and R. Stulz, 2004, “Are Daily Cross-Border Equity Flows Pushed or
Pulled?” Review of Economics and Statistics, Vol. 86, No. 3, pp. 642-657.
Hale, G., C. Candelaria, J. Caballero, and S. Borisov, 2011, “Global Banking Network and
Cross-Border Capital Flows,” mimeo, Federal Reserve Bank of San Francisco.
Haynes, S., 1988, “Identification of Interest Rates and International Capital Flows,” Review
of Economics and Statistics, Vol. 70(1), pp.103-111.
Hernandez, L., P. Mellado, and R. Valdes, 2001, “Determinants of Private Capital Flows in
the 1970s and 1990s: Is there Evidence of Contagion?” IMF Working Paper 01/64
(Washington DC: IMF).
IMF, 2011, World Economic Outlook: April 2011 (Washington DC: IMF).
IMF, 2010, The IMF-FSB Early Warning Exercise: Design and Methodological Toolkit,
(Washington DC: International Monetary Fund).
Kaminsky, G., R. Lyons, and S. Schmukler, 2001, “Mutual Fund Investment in Emerging
Markets: An Overview,” Policy Research Paper 2529 (Washington DC: World Bank).
Kass, G., 1980, “An Exploratory Technique for Investigating Large Quantities of Categorical
Data,” Applied Statistics, Vol. 29(2), pp. 119–127.
Kouri, P., and M. Porter, 1974, “International Capital Flows and Portfolio Equilibrium,”
Journal of Political Economy, Vol. 82(3), pp. 443-467.
Milesi-Ferretti, G. M., and A. Razin, 2000, “Current Account Reversals and Currency Crises,
Empirical Regularities,” in P. Krugman (ed.), Currency Crises (Chicago: University of
Chicago Press).
Milesi-Ferretti, G. M., and C. Tille, 2011, “The Great Retrenchment: International Capital
Flows During the Global Financial Crisis,” Economic Policy, Vol. 26(66), pp. 285-342.
Minoiu, C., and A. Reyes, 2011, “A Network Analysis of Global Banking: 19782009,” IMF
Working Paper 11/74, (Washington DC: International Monetary Fund).
Ostry, J., A. Ghosh, K. Habermeier, L. Laeven, M. Chamon, M. Qureshi, and A. Kokenyne,
“Managing Capital Inflows: What Tools to Use?” Staff Discussion Notes 11/06 (Washington
DC: International Monetary Fund).
Reinhart, C., and V. Reinhart, 2008, “Capital Flow Bonanzas: An Encompassing View of the
Past and Present,” NBER Working Paper 14321 (Cambridge: NBER).
Taylor, M., and L. Sarno, 1997, “Capital Flows to Developing Countries: Long- and ShortTerm Determinants,” World Bank Economic Review, Vo. 11(3), pp. 451-470.
27
Figure 1. Net Capital Flows to EMEs,
1980-2009 (in USD billions)
Figure 2. Surges of Net Capital Flows to EMEs,
1980-2009
60
Asia
Europe
Latin America & Caribb.
Middle East & Africa
In percent of GDP (Avg, right-axis)
550
12
6
40
400
8
250
2
20
4
100
0
-50
-2
1980
1985
1990
1995
2000
0
1980 1984 1988 1992 1996 2000 2004 2008
2005 2009
Avg. net financial flow to GDP in surges (in %)-right axis
Surge observations--threshold approach (in % of total obs.)
Source: IMF’s IFS database.
Source: Authors' estimates based on IFS.
Figure 3. Types of Surges, 19802009
(a) Number of surges based on type
(b) Average net capital flow in different types of surges
30
25
12
20
8
15
10
4
5
0
0
1980 1984 1988 1992 1996 2000 2004 2008
Asset-driven surges
Source: Authors’ estimates.
Liability-driven surges
1980-84 1985-89 1990-94 1995-99 2000-04 2005-09
Asset-driven
Asset-driven (avg.)
Liability-driven
Liability-driven (avg.)
28
Figure 4. Predicted Probabilities of Surge Occurrence
Predicted surge prob.
Predicted surge prob.
95 percent CI
Avg. real US interest rate
95 percent CI
0.3
0.2
Avg. S&P 500 index volatility
0.2
0.1
0.1
0.0
0
0.02
0.04
0.06
Real US interest rate
0.3
0.08
0.1
0.0
0
5
10
15
S&P 500 index volatility
0.8
Predicted surge prob.
95 percent CI
Avg. real domestic interest rate
20
Predicted surge prob.
95 percent CI
Avg. optimal CAB
0.6
0.2
0.4
0.1
0.2
0.0
0
0.02
0.04
0.06
0.08
0.1
0.0
-0.1
Real domestic interest rate
0.3
0.3
0.2
0.2
0.1
0.1
-0.05
0
0.05
Optimal current account balance to GDP
Predicted surge prob.
95 percent CI
Avg. real GDP growth rate
0.0
0
0.02
0.04
0.06
Real GDP growth rate
0.08
0.1
0.1
Predicted surge prob.
95 percent CI
Avg. capital acc. openness
0.0
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
Capital account openness index
Source: Authors’ estimates.
Notes: Predicted probabilities are based on the estimation results reported in Table 2 (column 10) holding all other variables fixed at mean value.
29
Figure 5. Predicted Probabilities of Asset- and Liability-Driven Surge Occurrence
0.30
Liability-driven surge
Asset-driven surge
Avg. real US interest rate
0.16
Liability-driven surge
Asset-driven surge
Avg. S&P500 index volatility
0.25
0.12
0.20
0.08
0.15
0.10
0.04
0.05
0.00
0
0.02
0.04
0.06
0.08
0.1
0.00
1
Real US interest rate
0.20
6
11
16
S&P 500 index volatility
Liability-driven surge
Asset-driven surge
Avg. real domestic interest rate
0.16
0.40
Liability-driven surge
Asset-driven surge
Avg. optimal current account balance
0.30
0.12
0.20
0.08
0.10
0.04
0.00
0.00
0
0.02
0.04
0.06
0.08
-0.1
0.1
Real domestic interest rate
0.25
0
0.05
0.1
Optimal current account balance to GDP
0.15
Liability-driven surge
Liability-driven surge
Asset-driven surge
Avg. capital account openness
Asset-driven surge
Avg. real GDP growth rate
0.20
-0.05
0.10
0.15
0.10
0.05
0.05
0.00
0
0.02
0.04
0.06
Real GDP growth rate
0.08
0.1
0.00
-2.0
-1.0
0.0
1.0
Capital account openness index
Source: Authors’ estimates.
Notes: Predicted probabilities for asset- and liability-driven surges are based on estimation results in Table 4 (panel b, cols. 5 and 10,
respectively) holding all other variables fixed at mean value.
2.0
Figure 6. Binary Recursive Trees for Surge Occurrence
(a) All surges
Surge: 232 (22 %)
No surge: 844 (78 %)
Y es
Optimal CA
Balance/GDP
< -1.3 pct.
Surge: 114 (15 %)
No surge: 632 (85 %)
Surge: 118 (36 %)
No surge: 212 (64 %)
S&P 500
index v olatility
< 20th pctile.
Y es
Surge: 38 (72 %)
No surge: 15 (28 %)
No
No
Surge:80 (29 %)
No surge: 197 (71 %)
(b) Types of surges
(i) Asset-driven surges
(ii) Liability-driven surges
Surge: 73 (08 %)
No surge: 844 (92 %)
Optimal
CA Balance/GDP
< -1.3 pct.
Yes
Surge: 43 (17 %)
No surge: 212 (83 %)
Optimal CA
Balance/GDP
< -9 pct.
Yes
Yes
Surge: 1 (20 %)
No surge: 4 (80 %)
Y es
No
Surge: 30 (05 %)
No surge: 632 (95 %)
Yes
Institutional
quality index
<70th pctile.
Surge: 36 (15 %)
Surge: 9 (02 %)
No surge: 207 (85 %) No surge: 417 (98 %)
Surge: 7 (58 %)
No surge: 5 (42 %)
Real GDP
Growth
< 6 pct.
No
Surge: 159 (16 %)
No surge: 844 (84 %)
No
Surge: 6 (86 %)
No surge: 1 (14 %)
S&P 500
index v olatility
< 20th pctile.
Surge: 61 (31 %)
No surge: 134 (69 %)
No
Surge: 21 (09 %)
No surge: 215 (91 %)
Y es
Optimal
No
CA balance/GDP
<-1.4 pct.
Surge: 27 (66 %)
No surge: 14 (34 %)
Surge: 34 (22 %)
No surge: 120 (78 %)
No
Surge: 98 (12 %)
No surge: 710 (88 %)
Y es
Optimal
No
CA Balance/GDP
<-0.2 pct.
Surge: 73 (18 %)
No surge: 332 (82 %)
Y es
Surge: 25 (06 %)
No surge: 378 (94 %)
Institutional
quality index
<99th pctile.
Surge: 21 (05 %)
No surge: 375 (95 %)
No
Surge: 4 (57 %)
No surge: 3 (43 %)
Source: Authors’ estimates.
Notes: Binary recursive trees have been constructed using the CHAID algorithm in SIPINA software (with the minimum size of nodes to split, and leaves specified as 10 and 5
observations, respectively; and the p-level for merging and splitting nodes specified as 0.05).
Table 1. Summary Statistics of Selected Variables
Surge
Observations
Mean
Min
Max Std dev.
Net capital flow s to GDP (in %)
232
10.56 ***
4.49
54.06
7.38
Real US interest rate (in %)
232
1.13 ***
-1.70
5.20
1.79
S&P 500 returns index volatility
232
7.47 ***
2.35
16.88
3.59
Real domestic interest rate (in %)
228
2.40
-23.93
41.65
5.92
REER overvaluation (in %)
232
0.81 *** -16.57
20.02
4.73
Optimal current account (in %)
232
-2.41 *** -21.56
Real GDP grow th rate
232
5.15 ***
Trade openness (in %)
232
84.93 ***
Reserves to GDP (in %)
232
16.50 ***
1.58
87.78
10.72
Real GDP per capita (Log)
232
8.08 ***
5.80
10.25
0.83
De facto exchange rate regime
232
1.89
1.00
3.00
1.89
Capital account openness index
232
0.63
-1.81
2.54
1.48
Financial interconnectedness
232
8.34 ***
1.00
15.00
3.28
Institutional quality index
232
0.67 ***
0.34
0.89
0.09
28.55
4.69
9.62
3.78
12.55
3.25
14.64 188.98
38.27
-9.26
Non-surge
Net capital flow s to GDP (in %)
844
1.00
-39.83
Real US interest rate (in %)
S&P 500 returns index volatility
844
1.54
-1.70
5.20
1.99
844
8.85
2.35
16.88
3.59
Real domestic interest rate (in %)
839
1.92
-38.00
30.31
6.61
REER overvaluation
844
-1.00
-45.00
59.00
7.63
Optimal current account (in %)
844
1.00
-11.00
20.23
3.37
Real GDP grow th (in %)
844
4.00
-15.00
25.65
4.23
Trade openness (in %)
844
67.33
12.10 220.41
37.43
Reserves to GDP (in %)
844
12.39
0.48 104.85
11.23
Real GDP per capita (Log)
844
7.73
5.59
10.20
0.94
De facto exchange rate regime
844
1.97
1.00
3.00
1.97
Capital account openness index
844
-0.08
-1.81
2.54
1.41
Financial interconnectedness
844
6.57
0.00
15.00
2.89
Institutional quality index
844
0.61
0.29
0.86
0.11
So urce: A utho rs' calculatio ns.
*Observatio ns restricted to the estimated sample as in Table 2. Real do mestic interest rate and real GDP
gro wth rate have been re-scaled using the fo rmula x/(1+x) if x≥0, and x/(1-x) if x<0 to transfo rm the o utliers. ***
indicates significant difference between the surge and no n-surge o bservatio ns at the 1percent level.
32
Table 2. Likelihood of Surge, 1980-2009
(1)
Real interest rate differential
[A] Constrained Model
(2)
(3)
(4)
(5)
0.581
1.347**
1.330**
1.392*
1.271*
(0.536)
(0.570)
(0.662)
(0.721)
(0.682)
Real US interest rate
(6)
[B] Unconstrained Model
(7)
(8)
(9)
(10)
-7.323** -12.133*** -8.893*** -9.593*** -11.918***
(3.499)
(2.930)
(2.941)
(3.015)
(3.033)
S&P500 index volatility
-0.038*** -0.057*** -0.052*** -0.046*** -0.034*** -0.040*** -0.064*** -0.058*** -0.052*** -0.040***
(0.012)
(0.012)
(0.012)
Commodity price index
1.030*** 0.912**
0.853**
0.655
(0.391)
(0.397)
(0.403)
(0.399)
Regional contagion
0.970*** 0.899***
0.461
0.301
0.297
(0.287)
(0.319)
(0.294)
(0.011)
(0.271)
(0.404)
(0.267)
(0.012)
(0.011)
Real domestic interest rate
REER overvaluation
(0.422)
(2.536)
(2.489)
(2.588)
(0.012)
(0.012)
1.094**
1.742***
(0.433)
(0.443)
(0.432)
0.260
0.082
0.029
(0.277)
(0.261)
(0.284)
(0.311)
(0.282)
0.818
1.202
1.351
1.469
1.484
(0.998)
(0.989)
(1.039)
(1.129)
(0.994)
-0.105
-1.157**
-1.134*
-1.114
-0.912
(0.537)
(0.640)
(0.742)
(0.696)
-10.112*** -9.931*** -9.887*** -9.880***
(0.424)
(0.012)
1.252***
0.763*** 0.542**
(0.497)
Optimal current account/GDP
(0.012)
1.124*** 1.411*** 1.452***
-11.106*** -10.683*** -10.688*** -10.984***
(2.626)
(2.764)
(2.648)
(2.752)
(2.838)
6.071***
4.804***
4.789***
Real GDP grow th
6.601*** 5.444*** 5.551***
(1.669)
(1.610)
(1.634)
(1.736)
(1.690)
(1.735)
Capital account openness
0.098**
0.104**
0.130**
0.079*
0.083*
0.114**
(0.048)
(0.052)
(0.047)
(0.047)
Financial interconnectedness
Exchange rate regime
(0.048)
(0.051)
0.074*** 0.076***
0.076***
0.081***
(0.016)
(0.016)
(0.016)
(0.015)
-0.118
-0.081
-0.134
-0.094
(0.093)
(0.089)
(0.094)
Institutional quality index
2.562***
Default onset
(0.749)
(0.739)
-0.291
-0.199
(0.367)
Real GDP per capita (log)
(0.344)
-0.329***
-0.399***
(0.108)
Observations
Regional dummies
Pseudo-R2
Percent correctly predicted
Sensitivity
Specificity
1,076
Yes
0.079
79.18
13.79
97.16
1,076
Yes
0.129
79.46
16.38
96.80
1,076
Yes
0.157
80.30
21.12
96.56
1,076
Yes
0.177
81.13
26.29
96.21
(0.088)
2.638***
1,076
Yes
0.198
81.23
27.16
96.09
(0.111)
1,076
Yes
0.085
79.46
14.22
97.39
1,076
Yes
0.143
80.20
20.26
96.68
1,076
Yes
0.164
80.39
22.84
96.21
1,076
Yes
0.184
81.23
25.86
96.45
1,076
Yes
0.209
81.51
28.88
95.97
So urce: A utho rs' estimates.
No tes: Dependent variable is a binary variable equal to 1if a surge o ccurs and 0 o therwise. Co nstrained mo del refers to the specificatio n where real interest rate
differential between co untry i and the US (real do mestic interest rate-real US interest rate-REER o vervaluatio n) is included. A ll regressio ns are estimated using a
pro bit mo del, with clustered standard erro rs (at the co untry level) repo rted in parentheses. Co nstant and regio n-specific effects are included in all specificatio ns.
***,**,* indicate significance at 1, 5, and 10 percent levels, respectively. Sensitivity (specificity) gives the fractio n o f surge (no -surge) o bservatio ns that are
co rrectly specified. A ll variables except fo r glo bal facto rs (real US interest rate, S&P 500 index vo latility, and co mmo dity price index), regio nal co ntagio n, and
financial interco nnectedness are lagged o ne perio d.
Table 3. Magnitude of Surge, 1980-2009
(1)
Real interest rate differential
[A] Constrained Model
(2)
(3)
(4)
0.024 0.040 0.006
0.077
(0.057) (0.060) (0.059) (0.054)
Real US interest rate
S&P 500 index volatility
Commodity price index
Regional contagion
0.000
(0.002)
0.063**
(0.031)
-0.149*
(0.085)
-0.000
(0.002)
0.055*
(0.030)
-0.129
(0.085)
-0.000
(0.002)
0.044
(0.035)
-0.113
(0.094)
Real domestic interest rate
REER overvaluation
Optimal current account/GDP
-0.295* -0.203
(0.158) (0.152)
0.001
(0.260)
0.010**
(0.004)
Real GDP grow th
Capital account openness
Financial interconnectedness
Exchange rate regime
Institutional quality index
Default onset
Real GDP per capita (log)
Observations
Regional dummies
R2
232
Yes
0.165
232
Yes
0.187
232
Yes
0.221
(5)
(6)
[B] Unconstrained Model
(7)
(8)
(9)
(10)
0.080
(0.060)
-0.979*** -1.036*** -0.944*** -0.885*** -0.899***
(0.320) (0.326) (0.317) (0.297) (0.298)
-0.001 -0.001
-0.001 -0.001 -0.001 -0.002 -0.002
(0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.001)
0.050
0.046
0.060* 0.050
0.046 0.054* 0.052
(0.032) (0.040) (0.030) (0.031) (0.033) (0.030) (0.037)
-0.146* -0.152* -0.248*** -0.229*** -0.207** -0.225** -0.232***
(0.087) (0.088) (0.085) (0.083) (0.087) (0.085) (0.086)
-0.168* -0.160* -0.193** -0.089 -0.100
(0.096) (0.094) (0.091) (0.062) (0.068)
-0.198** -0.226** -0.214** -0.238** -0.253***
(0.089) (0.094) (0.099) (0.095) (0.093)
-0.238 -0.225
-0.347** -0.274* -0.294** -0.284*
(0.154) (0.168)
(0.145) (0.141) (0.143) (0.158)
-0.151 -0.181
-0.178 -0.282 -0.286
(0.251) (0.260)
(0.247) (0.245) (0.253)
0.008*
0.006
0.008** 0.006
0.004
(0.004) (0.004)
(0.004) (0.004) (0.004)
-0.001 -0.001
-0.001 -0.001
(0.002) (0.002)
(0.002) (0.002)
-0.042*** -0.042***
-0.037*** -0.037***
(0.013) (0.014)
(0.011) (0.012)
0.003
0.045
(0.097)
(0.087)
-0.028
0.008
(0.033)
(0.025)
0.007
0.006
(0.022)
(0.021)
232
Yes
0.319
232
Yes
0.323
232
Yes
0.248
232
Yes
0.278
232
Yes
0.304
232
Yes
0.378
232
Yes
0.384
So urce: A utho rs' estimates.
No tes: Dependent variable is net capital flo w to GDP if a surge o ccurs. Co nstrained mo del refers to the specificatio n where real interest rate
differential between co untry i and the US (real do mestic interest rate-real US interest rate-REER o vervaluatio n) is included. A ll regressio ns
estimated using po o led OLS. Co nstant, and regio nal specific effects are included in all specificatio ns. A ll variables except fo r glo bal facto rs (real
US interest rate, S&P 500 index vo latility, and co mmo dity price index), regio nal co ntagio n, and financial interco nnectedness are lagged o ne perio d.
Clustered standard erro rs (at the co untry level) repo rted in parentheses. ***,**,* indicate significance at 1, 5, and 10 percent levels, respectively.
34
Table 4. Likelihood of Surge: by Surge Type, 1980-2009
Asset-driven surge
(1)
Real interest rate differential -0.252
(0.530)
S&P 500 index volatility
-0.016
(0.014)
Commodity price index
2.219***
(0.476)
Regional contagion
0.433
(0.344)
Optimal current account/GDP
Real GDP grow th
Capital account openness
Financial interconnectedness
Exchange rate regime
Institutional quality index
Default onset
Real GDP per capita (log)
Observations
Regional dummies
Pseudo-R2
917
Yes
0.083
(2)
(3)
Liability-driven surge
(4)
(5)
(6)
(7)
[A] Constrained Model
0.749
0.450
0.425
0.253
0.870
1.488**
(0.686)
(0.813)
(0.840)
(0.672)
(0.554) (0.608)
-0.045*** -0.039** -0.038** -0.023
-0.050*** -0.066***
(0.017)
(0.017)
(0.016)
(0.018)
(0.014) (0.015)
2.366*** 2.307*** 2.261*** 2.648***
0.245
0.113
(0.592)
(0.592)
(0.601)
(0.639)
(0.414) (0.419)
0.303
-0.282
-0.315
-0.303
1.142*** 1.064***
(0.359)
(0.441)
(0.443)
(0.417)
(0.278) (0.269)
-13.696*** -13.685*** -13.682*** -13.690***
-8.101***
(2.621)
(2.642)
(2.643)
(2.553)
(2.731)
5.348*** 5.178*** 4.892***
(1.906)
(1.919)
(1.848)
0.179*** 0.181*** 0.198***
(0.060)
(0.059)
(0.063)
0.018
0.016
(0.021)
(0.022)
0.008
0.052
(0.115)
(0.116)
2.922***
(0.895)
-0.326
(0.436)
-0.261**
(0.112)
917
917
917
917
1,003
1,003
Yes
Yes
Yes
Yes
Yes
Yes
0.174
0.215
0.216
0.239
0.0875
0.119
(8)
(9)
(10)
1.664**
(0.683)
-0.063***
(0.014)
0.021
(0.422)
0.728**
(0.289)
-8.209***
(2.667)
6.682***
(1.820)
0.046
(0.043)
1.820**
(0.772)
-0.054***
(0.014)
-0.285
(0.444)
0.529*
(0.308)
-8.232***
(2.904)
5.168***
(1.746)
0.054
(0.046)
0.096***
(0.020)
-0.160*
(0.091)
1,003
Yes
0.143
1,003
Yes
0.176
1.691**
(0.749)
-0.046***
(0.014)
0.174
(0.440)
0.484
(0.302)
-8.301***
(2.950)
5.541***
(1.809)
0.088*
(0.049)
0.100***
(0.020)
-0.121
(0.089)
2.207***
(0.741)
-0.487
(0.473)
-0.348***
(0.116)
1,003
Yes
0.196
[B] Unconstrained Model
Real US interest rate
S&P 500 index volatility
Commodity price index
Regional contagion
Real domestic interest rate
REER overvaluation
-7.775*
(4.547)
-0.014
(0.015)
2.680***
(0.522)
0.218
(0.358)
1.129
(0.923)
1.673***
(0.636)
Optimal current account/GDP
-15.485***
(4.009)
-0.051***
(0.018)
3.104***
(0.624)
-0.206
(0.352)
1.975**
(0.987)
0.616
(0.706)
-15.517***
(2.724)
-10.815***
(4.093)
-0.045**
(0.019)
2.876***
(0.628)
-0.542
(0.418)
1.822*
(1.030)
0.954
(0.860)
-14.955***
(2.695)
4.968**
(1.983)
0.147**
(0.062)
-11.068***
(4.123)
-0.043**
(0.018)
2.821***
(0.630)
-0.582
(0.421)
1.817*
(1.034)
0.998
(0.900)
-14.933***
(2.715)
4.708**
(2.000)
0.146**
(0.060)
0.022
(0.021)
-0.026
(0.107)
917
Yes
0.206
917
Yes
0.233
917
Yes
0.234
Real GDP grow th
Capital account openness
Financial interconnectedness
Exchange rate regime
Institutional quality index
Default onset
Real GDP per capita (log)
Observations
Regional dummies
Pseudo-R2
917
Yes
0.101
-12.555***
(4.692)
-0.028
(0.020)
3.367***
(0.690)
-0.575
(0.392)
1.879*
(0.966)
1.079
(0.691)
-15.323***
(2.693)
4.213**
(1.921)
0.177***
(0.065)
0.027
(0.023)
0.026
(0.109)
2.940***
(0.890)
-0.256
(0.364)
-0.360***
(0.118)
917
Yes
0.258
-5.810*
(3.278)
-0.053***
(0.014)
0.476
(0.442)
0.996***
(0.287)
0.401
(1.029)
-1.125**
(0.567)
-9.545***
(2.936)
-0.073***
(0.015)
0.467
(0.442)
0.817***
(0.269)
0.695
(1.034)
-2.035***
(0.613)
-8.975***
(2.898)
-7.163**
(2.967)
-0.069***
(0.015)
0.268
(0.465)
0.589**
(0.291)
1.053
(1.077)
-2.224***
(0.689)
-8.882***
(2.777)
6.172***
(1.878)
0.036
(0.043)
-7.702**
(3.090)
-0.060***
(0.015)
0.003
(0.489)
0.379
(0.306)
1.236
(1.242)
-2.326***
(0.788)
-8.944***
(3.022)
4.563**
(1.833)
0.043
(0.047)
0.096***
(0.021)
-0.164*
(0.089)
1,003
Yes
0.0914
1,003
Yes
0.129
1,003
Yes
0.148
1,003
Yes
0.181
-10.288***
(3.074)
-0.053***
(0.015)
0.660
(0.475)
0.272
(0.293)
1.228
(1.115)
-2.016**
(0.826)
-9.248***
(3.105)
4.797**
(1.919)
0.076
(0.049)
0.102***
(0.019)
-0.125
(0.087)
2.333***
(0.740)
-0.401
(0.467)
-0.397***
(0.121)
1,003
Yes
0.205
So urce: A utho rs' estimates.
No tes: Dependent variable is a binary variable (=1if a surge o ccurs; 0 o therwise). A sset- (liability-) driven surge is defined as the surge when change in residents'
assets is larger than the change in their liabilities (assets). Regressio ns estimated using pro bit mo del, with clustered standard erro rs (at the co untry level) repo rted
in parentheses. A ll variables except fo r real US interest rate, S&P 500 index vo latility, co mmo dity price index, regio nal co ntagio n and financial interco nnectedness
are lagged o ne perio d. Co nstant and regio n-specific effects are included in all specificatio ns. ***,**,* indicate significance at 1, 5, and 10 percent levels, respectively.
35
Table 5. Magnitude of Surge: by Surge Type, 1980-2009
(1)
Real interest rate differential
S&P500 index volatility
Commodity price index
Regional contagion
Asset-driven surge
(2)
(3)
(4)
0.046
(0.013)
-0.000
(0.001)
0.089
(0.053)
0.027
(0.045)
0.047
(0.076)
-0.000
(0.001)
0.089
(0.054)
0.028
(0.043)
-0.001
(0.190)
0.005
(0.062)
-0.001
(0.001)
0.069
(0.054)
0.032
(0.056)
0.034
(0.178)
-0.100
(0.156)
0.012**
(0.005)
73
Yes
0.263
73
Yes
0.263
73
Yes
0.330
-0.506
(0.371)
-0.001
(0.001)
0.066
(0.052)
-0.035
(0.046)
-0.086
(0.117)
-0.107
(0.078)
-0.506
(0.373)
-0.001
(0.001)
0.066
(0.053)
-0.034
(0.043)
-0.086
(0.118)
-0.107
(0.078)
-0.009
(0.174)
-0.479
(0.347)
-0.001
(0.001)
0.049
(0.050)
-0.019
(0.053)
-0.127
(0.099)
-0.095
(0.081)
0.019
(0.152)
-0.218
(0.164)
0.011**
(0.005)
73
Yes
0.304
73
Yes
0.304
73
Yes
0.376
Optimal current account/GDP
Real GDP grow th
Capital account openness
Financial interconnectedness
Exchange rate regime
Institutional quality index
Default onset
Real GDP per capita (log)
Observations
Regional dummies
R2
Real US interest rate
S&P500 index volatility
Commodity price index
Regional contagion
Real domestic interest rate
REER overvaluation
Optimal current account/GDP
Real GDP grow th
Capital account openness
Financial interconnectedness
Exchange rate regime
Institutional quality index
Default onset
Real GDP per capita (log)
Observations
Regional dummies
R2
Liability-driven surge
(5)
(6)
(7)
(8)
(9)
[A] Constrained Model
0.050
(0.065)
-0.002
(0.001)
0.085
(0.060)
0.008
(0.055)
0.016
(0.165)
-0.105
(0.149)
0.012**
(0.005)
-0.002
(0.003)
-0.023*
(0.012)
-0.003
-0.002
0.024
(0.089)
(0.069) (0.076)
-0.002
0.001
0.000
(0.002)
(0.002) (0.002)
0.053
0.087*
0.065
(0.057)
(0.052) (0.049)
-0.035
-0.280** -0.268**
(0.066)
(0.137) (0.123)
0.044
-0.501***
(0.181)
(0.182)
-0.093
(0.242)
0.010*
(0.006)
-0.004
(0.004)
-0.024*
(0.014)
-0.025
(0.107)
-0.033
(0.039)
0.020
(0.029)
73
73
159
159
Yes
Yes
Yes
Yes
0.382
0.405
0.178
0.231
[B] Unconstrained Model
-0.403
-0.373
-1.229*** -1.327***
(0.296) (0.368)
(0.380) (0.391)
-0.002* -0.003*
0.000
-0.001
(0.001) (0.001)
(0.002) (0.002)
0.066
0.031
0.102**
0.078
(0.055) (0.058)
(0.050) (0.049)
-0.029
-0.073
-0.381** -0.375***
(0.053) (0.063)
(0.142) (0.122)
-0.071
-0.126
-0.203* -0.186*
(0.087) (0.118)
(0.113) (0.110)
-0.122
-0.097
-0.196* -0.240**
(0.083) (0.091)
(0.106) (0.102)
0.006
0.031
-0.558***
(0.152) (0.171)
(0.171)
-0.205
-0.139
(0.157) (0.236)
0.011**
0.009
(0.005) (0.006)
-0.002
-0.004
(0.003) (0.004)
-0.019
-0.021
(0.011) (0.014)
-0.013
(0.117)
0.005
(0.036)
0.022
(0.030)
73
73
159
159
Yes
Yes
Yes
Yes
0.411
0.437
0.269
0.334
(10)
-0.003
0.082
0.077
(0.078) (0.066) (0.079)
0.001
0.000
0.001
(0.003) (0.002) (0.002)
0.073
0.062
0.065
(0.059) (0.053) (0.058)
-0.239* -0.259** -0.255*
(0.126) (0.125) (0.129)
-0.391** -0.443** -0.428**
(0.186) (0.171) (0.191)
-0.054
-0.282
-0.293
(0.377) (0.350) (0.345)
0.009*
0.005
0.004
(0.004) (0.004) (0.005)
-0.000
-0.001
(0.002) (0.002)
-0.050*** -0.051***
(0.016) (0.016)
0.039
(0.122)
0.016
(0.037)
0.006
(0.025)
159
159
159
Yes
Yes
Yes
0.253
0.364
0.368
-1.275***
(0.397)
-0.000
(0.002)
0.090
(0.057)
-0.356***
(0.124)
-0.210*
(0.114)
-0.225**
(0.106)
-0.488**
(0.185)
-0.221
(0.357)
0.006
(0.004)
-1.195***
(0.383)
-0.001
(0.002)
0.083
(0.052)
-0.358***
(0.127)
-0.079
(0.075)
-0.240**
(0.104)
-0.519***
(0.158)
-0.384
(0.345)
0.003
(0.004)
-0.001
(0.002)
-0.044***
(0.014)
159
Yes
0.347
159
Yes
0.431
-1.254***
(0.380)
-0.001
(0.002)
0.098*
(0.057)
-0.359***
(0.132)
-0.094
(0.081)
-0.251**
(0.108)
-0.506***
(0.184)
-0.383
(0.336)
0.001
(0.004)
-0.001
(0.002)
-0.044***
(0.014)
0.100
(0.108)
0.019
(0.032)
0.001
(0.023)
159
Yes
0.440
So urce: A utho rs' estimates.
No tes: Dependent variable is net capital flo w to GDP if a surge o ccurs. A sset- (liability-) driven surge is the surge when change in residents' assets- (liabilities-) is
larger than the change in residents' liabilities (assets). A ll variables except fo r real US interest rate, S&P 500 index vo latility, co mmo dity price index, regio nal
co ntagio n and financial interco nnectedness are lagged o ne perio d. Co nstant, and regio n specific effects are included. A ll regressio ns estimated using OLS.
Clustered standard erro rs (at the co untry level) repo rted in parentheses. ***,**,* indicate significance at 1, 5, and 10 percent levels, respectively.
Table 6. Likelihood of Surge: Sensitivity Analysis
Surge definitions
Real US interest rate
Extended
Cluster
(1)
(2)
-14.295***
(3.425)
S&P500 index/RAI/VXO
-0.045***
(0.013)
Commodity price index
1.479***
(0.493)
Regional contagion
0.321
(0.330)
Real domestic interest rate
2.279**
(1.030)
REER overvaluation
-1.198**
(0.510)
Optimal current account/GDP -12.441***
(3.049)
Real GDP grow th
4.037**
(1.652)
Capital account openness
0.134**
(0.061)
Financial interconnectedness 0.074***
(0.017)
Exchange rate regime
-0.140
(0.102)
Institutional quality index
2.395***
(0.705)
Default onset
-0.514*
(0.290)
Real GDP per capita (log)
-0.369***
(0.121)
Observations
1,076
Regional dummies
Yes
Pseudo-R2
0.221
-13.235***
(3.293)
-0.036**
(0.016)
1.510***
(0.450)
0.516**
(0.252)
2.354**
(0.993)
-1.141
(0.712)
-9.902***
(2.758)
5.145***
(1.821)
0.028
(0.040)
0.092***
(0.014)
-0.088
(0.089)
1.603**
(0.674)
-0.473
(0.381)
-0.229**
(0.095)
1,076
Yes
0.192
Alternate regressors
Real US
RAI
VXO
10yr yield
(3)
(4)
(5)
-6.150**
(3.105)
-0.028**
(0.012)
1.343***
(0.409)
0.186
(0.279)
1.224
(0.964)
-0.877
(0.708)
-10.603***
(2.754)
5.050***
(1.735)
0.115**
(0.052)
0.080***
(0.016)
-0.085
(0.090)
2.311***
(0.750)
-0.222
(0.361)
-0.365***
(0.110)
1,076
Yes
0.202
Additional regressors
Trade Reserves Stock market Return Private sector
openness to GDP capitalization on equity credit/GDP
(6)
(7)
(8)
(9)
(10)
-8.398** -11.299*** -10.070***
(3.802) (4.108)
(2.884)
-0.009 -0.041***
0.165*
(0.095) (0.008)
(0.013)
1.845*** 1.718*** 1.725***
(0.463) (0.478)
(0.444)
0.138
0.256
-0.017
(0.305) (0.307)
(0.279)
1.498
1.648
1.813*
(1.156) (1.227)
(0.980)
-0.865
-1.011
-0.401
(0.711) (0.664)
(0.692)
-11.356***-10.511*** -11.052***
(3.047) (2.987)
(2.571)
4.893*** 5.071*** 4.473**
(1.884) (1.817)
(1.776)
0.077
0.093*
0.094*
(0.051) (0.050)
(0.052)
0.092*** 0.092*** 0.102***
(0.016) (0.016)
(0.014)
-0.129
-0.137
-0.036
(0.087) (0.087)
(0.096)
2.545*** 2.691*** 1.853**
(0.811) (0.791)
(0.844)
-0.295
-0.324
-0.238
(0.503) (0.509)
(0.344)
-0.466*** -0.448*** -0.429***
(0.116) (0.116)
(0.105)
998
944
1,076
Yes
Yes
Yes
0.216
0.207
0.223
-10.600***
(2.764)
-0.039***
(0.012)
1.699***
(0.432)
0.035
(0.278)
1.597
(1.008)
-0.758
(0.706)
-11.017***
(2.881)
4.853***
(1.735)
0.104**
(0.050)
0.084***
(0.016)
-0.079
(0.089)
2.397***
(0.738)
-0.176
(0.345)
-0.427***
(0.111)
1,076
Yes
0.212
-13.098*** -10.649** -11.910***
(3.737)
(4.583)
(2.922)
-0.053***
-0.033*
-0.040***
(0.016)
(0.018)
(0.012)
1.624***
1.808***
1.752***
(0.573)
(0.540)
(0.428)
-0.090
-0.291
0.019
(0.325)
(0.387)
(0.286)
1.241
1.517
1.507
(1.314)
(1.313)
(1.017)
-0.919
-1.633**
-1.619*
(0.792)
(0.853)
(0.683)
-13.980*** -11.039*** -10.940***
(2.792)
(3.186)
(2.830)
4.849**
5.547**
4.862***
(2.348)
(2.215)
(1.768)
0.082
0.051
0.112**
(0.055)
(0.050)
(0.051)
0.093***
0.113***
0.081***
(0.016)
(0.016)
(0.015)
-0.087
-0.160
-0.200**
(0.102)
(0.091)
(0.091)
3.334***
2.369**
2.580***
(0.929)
(0.930)
(0.765)
-0.058
-0.213
(0.567)
(0.342)
-0.462***
-0.432***
-0.411***
(0.148)
(0.118)
(0.128)
825
766
1,066
Yes
Yes
Yes
0.248
0.217
0.208
Trade
links
(11)
-12.457***
(3.385)
-0.041***
(0.012)
1.708***
(0.454)
0.012
(0.280)
1.609
(1.015)
-0.944
(0.708)
-10.963***
(2.947)
4.787***
(1.854)
0.121**
(0.052)
0.083***
(0.017)
-0.084
(0.094)
2.604***
(0.833)
-0.197
(0.341)
-0.415***
(0.138)
1,036
Yes
0.208
Fixed
effects
(12)
Sam ple Estim ation
1990- Complementary
2009
Log-Log
(13)
(14)
-15.919*** -12.009***
(4.086) (4.446)
-0.043*** -0.043***
(0.014) (0.017)
2.037*** 1.674***
(0.525) (0.484)
0.232
-0.325
(0.317) (0.345)
1.617
1.021
(1.058) (1.280)
-1.246 -1.887**
(0.801) (0.786)
-13.407*** -12.211***
(2.659) (3.401)
5.401** 4.788**
(2.248) (1.903)
0.056
0.133*
(0.080) (0.049)
0.141*** 0.102***
(0.022) (0.016)
-0.228 -0.181**
(0.149) (0.089)
2.967*** 2.240**
(1.127) (0.884)
-0.145
-0.226
(0.355) (0.540)
-0.708*** -0.438***
(0.225) (0.116)
1,036
832
Yes
Yes
0.304
0.207
-17.543***
(4.820)
-0.069***
(0.017)
2.261***
(0.640)
-0.297
(0.409)
2.335
(1.500)
-0.798
(0.989)
-16.601***
(3.640)
6.949***
(2.395)
0.163**
(0.072)
0.132***
(0.022)
-0.071
(0.128)
4.280***
(1.195)
-0.674
(0.655)
-0.645***
(0.167)
1,076
Yes
So urce: A utho rs' estimates.
No tes: Dependent variable is a binary variable (=1if a surge o ccurs; 0 o therwise). A ll regressio ns (except fo r co mplemetary lo g-lo g regressio n) are estimated using pro bit estimatio n metho d. Clustered standard erro rs (at the co untry
level) are repo rted in parentheses. A ll variables except fo r real US interest rate, S&P 500 index vo latility, co mmo dity price index, regio nal co ntagio n, and financial interco nnectedness are lagged o ne perio d. Co nstant and regio n
specific effects are included in all specificatio ns. ***,**,* indicate significance at 1, 5, and 10 percent levels, respectively. Extended=Surges identified using a o ne-year windo w (i.e., including the year befo re and after the surgeif the net
capital flo w is po sitive); Cluster=Surges identified using the cluster appro ach; Real US 10yr yield=Including the real US 10 yr go vernment bo nd yield instead o f the real US 3-mo nth T-bill rate; RA I=including the (lo g o f) Credit Suisse
glo bal risk appetite index (RA I) instead o f the S&P 500 index vo latility measure; VXO=Including the VXO index instead o f the S&P 500 vo latility index; Trade o penness=Including trade to GDP ratio in the specificatio n; Reserves to
GDP =Including the sto ck o f fo reign reserves to GDP ratio in the specificatio n; Sto ck market capitalizatio n=Including sto ck market capitalizatio n in the specificatio n; Return o n equity=Including banks's return o n equity in the
specificatio n; P rivate secto r credit/GDP =Including private secto r credit to GDP ratio in the specificatio n; Trade links=Including trade links to measure co ntagio n effects in the specificatio n; Fixed effects=Including co untry fixed
37
Table 7. Magnitude of Surge: Sensitivity Analysis
Surge definitions
Extended
(1)
-0.735**
(0.278)
S&P 500 index/RAI/VXO
-0.001
(0.001)
Commodity price index
0.030**
(0.015)
Regional contagion
-0.296**
(0.117)
Real domestic interest rate
-0.044
(0.044)
REER overvaluation
-0.150***
(0.054)
Optimal current account/GDP -0.315**
(0.127)
Real GDP grow th
-0.116
(0.168)
Capital account openness
0.000
(0.002)
Financial interconnectedness 0.004
(0.003)
Exchange rate regime
-0.026***
(0.008)
Institutional quality index
0.071
(0.064)
Default onset
0.013
(0.014)
Real GDP per capita (log)
-0.003
(0.015)
Observations
365
Regional dummies
Yes
R-squared
0.294
Real US interest rate
Cluster
(2)
-0.663**
(0.293)
-0.001
(0.001)
0.034
(0.030)
0.005
(0.024)
-0.075
(0.055)
-0.189**
(0.080)
-0.358**
(0.135)
-0.031
(0.193)
0.006
(0.004)
-0.002
(0.002)
-0.030**
(0.011)
0.081
(0.083)
0.118
(0.071)
0.001
(0.015)
278
Yes
0.353
Alternate regressors
Real US
RAI
VXO
10yr yield
(3)
-0.441
(0.277)
-0.001
(0.001)
0.023
(0.036)
-0.198**
(0.088)
-0.131*
(0.074)
-0.247**
(0.096)
-0.278*
(0.157)
-0.267
(0.256)
-0.001
(0.002)
0.005
(0.004)
-0.037***
(0.012)
0.004
(0.090)
0.002
(0.026)
0.009
(0.021)
232
Yes
0.362
Additional regressors
Trade
Reserves Stock market Return on Private sector
openness
to GDP capitalization equity
credit/GDP
(4)
(5)
(6)
-0.837*** -0.800** -0.537*
(0.311)
(0.320)
(0.274)
-0.005
-0.002
-0.001*
(0.008)
(0.001)
(0.001)
0.057
0.055
0.053
(0.041)
(0.041)
(0.035)
-0.253** -0.286*** -0.183**
(0.097)
(0.098)
(0.082)
-0.036
-0.143** -0.143**
(0.068)
(0.070)
(0.068)
-0.283*** -0.295*** -0.185**
(0.104)
(0.103)
(0.090)
-0.269
-0.313*
-0.299*
(0.162)
(0.162)
(0.163)
-0.300
-0.361
-0.273
(0.251)
(0.256)
(0.259)
0.005
-0.001
0.000
(0.004)
(0.002)
(0.002)
-0.001
0.003
0.002
(0.002)
(0.004)
(0.004)
-0.036*** -0.036*** -0.030***
(0.011)
(0.011)
(0.010)
0.047
0.045
-0.008
(0.095)
(0.104)
(0.093)
0.016
0.002
-0.011
(0.024)
(0.027)
(0.027)
0.006
0.009
0.001
(0.022)
(0.022)
(0.020)
220
216
232
Yes
Yes
Yes
0.392
0.411
0.428
(7)
-0.415
(0.325)
-0.001
(0.001)
0.030
(0.034)
-0.202**
(0.093)
-0.040
(0.067)
-0.187*
(0.101)
-0.192
(0.153)
-0.247
(0.260)
-0.001
(0.002)
0.002
(0.004)
-0.032***
(0.008)
-0.022
(0.083)
0.008
(0.028)
-0.001
(0.014)
232
Yes
0.485
(8)
-0.820**
(0.314)
-0.002
(0.002)
0.045
(0.044)
-0.311***
(0.092)
-0.190**
(0.088)
-0.283**
(0.110)
-0.293
(0.195)
-0.226
(0.270)
-0.002
(0.002)
0.003
(0.004)
-0.037***
(0.013)
-0.003
(0.113)
0.000
(0.000)
0.016
(0.025)
196
Yes
0.415
(9)
-0.773*
(0.385)
-0.002
(0.002)
0.043
(0.041)
-0.236**
(0.108)
-0.129
(0.089)
-0.252**
(0.099)
-0.264
(0.160)
-0.213
(0.292)
-0.001
(0.002)
0.005
(0.004)
-0.036***
(0.012)
0.056
(0.117)
0.018
(0.026)
0.007
(0.025)
198
Yes
0.384
(10)
-0.875***
(0.298)
-0.002
(0.001)
0.051
(0.037)
-0.225**
(0.088)
-0.083
(0.068)
-0.248***
(0.092)
-0.275*
(0.157)
-0.297
(0.256)
-0.001
(0.002)
0.005
(0.004)
-0.036***
(0.012)
0.039
(0.086)
0.001
(0.027)
0.003
(0.023)
232
Yes
0.385
Trade
links
(11)
-0.863***
(0.316)
-0.001
(0.001)
0.048
(0.043)
-0.233***
(0.085)
-0.074
(0.064)
-0.258***
(0.090)
-0.350**
(0.164)
-0.265
(0.256)
-0.002
(0.002)
0.003
(0.004)
-0.035***
(0.012)
0.031
(0.110)
0.002
(0.025)
0.010
(0.026)
224
Yes
0.399
Sam ple
Fixed
1990-2009
effects
(12)
(13)
-0.071 -0.797***
(0.198)
(0.276)
-0.003*** -0.001
(0.001)
(0.001)
0.067*
0.067**
(0.036)
(0.026)
-0.032 -0.252***
(0.062)
(0.083)
0.008
-0.056
(0.063)
(0.049)
-0.175** -0.162**
(0.081)
(0.062)
-0.143
-0.356**
(0.153)
(0.145)
-0.463
-0.138
(0.326)
(0.195)
-0.001
0.001
(0.001)
(0.002)
0.000
0.004
(0.008)
(0.003)
-0.017* -0.026***
(0.009)
(0.008)
-0.083
0.066
(0.071)
(0.073)
-0.058*
0.024
(0.031)
(0.015)
0.009
-0.003
(0.013)
(0.016)
232
338
Yes
Yes
0.801
0.298
So urce: A utho rs' estimates.
No tes: Dependent variable is net capital flo w to GDP co nditio nal o n surge o ccurrence. A ll regressio ns are estimated using OLS, with clustered standard erro rs (at the co untry level) repo rted in parentheses. A ll
variables except fo r real US interest rate, S&P 500 index vo latility, co mmo dity price index, regio nal co ntagio n, and financial interco nnectedness are lagged o ne perio d. Co nstant and regio n specific effects are included
in all specificatio ns. ***,**,* indicate significance at 1, 5, and 10 percent levels, respectively. Extended=Surges identified using a o ne-year windo w (i.e., including the year befo re and after the surge if the net capital flo w is
po sitive); Cluster=Surges identified using the cluster appro ach; Real US 10yr yield=Including the real US 10 yr go vernment bo nd yield instead o f the real US 3-mo nth T-bill rate; RA I=Including the (lo g o f) Credit Suisse
glo bal risk appetite index (RA I) instead o f the S&P 500 index vo latility measure; VXO=Including the VXO index instead o f the S&P 500 index vo latility measure; Trade o penness=Including trade to GDP ratio in the
specificatio n; Reserves to GDP =Including sto ck o f fo reign reserves to GDP ratio in the specificatio n; Sto ck market capitalizatio n=Including sto ck market capitalizatio n in the specificatio n; Return o n
equity=Including banks's return o n equity in the specificatio n; P rivate secto r credit/GDP =Including private secto r credit to GDP ratio in the specificatio n; Trade links=Including trade links to measure co ntagio n
effects in the specificatio n; Fixed effects=Including co untry fixed effects in the specificatio n.
APPENDIX A: THE INTERTEMPORAL OPTIMIZING MODEL OF THE CURRENT ACCOUNT
Capital flows to EMEs should correspond to their external financing needs; to proxy for the
latter, we use an intertemporal optimizing model of the current account following Ghosh (1995).
If the country can borrow (or lend) freely in the world capital markets, then consumption need
not depend on the current realization of “national cash flow” (output, net of investment and
government consumption) but rather on the annuity value of its entire present value:
(Q  I  G ) 
1 
 r 
ct*    bt 
Et t  j t  j j t  j 

(1  r ) j 0
(1  r )
 

(8)
where θ is a constant of proportionality reflecting consumption tilting given the country’s
subjective discount rate and the world interest rate, c* reflects consumption, Q is GDP, I is
investment, and G is government consumption. The assumption that the economy is small in the
world capital markets implies Fisherian separability: investment is undertaken until the marginal
product of capital equals the world interest rate. Thus investment and output can be taken as
given when making the consumption decision. The consumption-smoothing component of the
current account (i.e., abstracting from consumption-tilting) is given by:
CAt*  Yt  I t  Gt   Ct*
(9)
where Y is GNP. Substituting for consumption, yields (after some manipulation):

CAt*  
j 1
Et (Qt  j  I t  j  Gt  j )
(1  r ) j
(10)
The expression for the current account (10) is fundamental to the intertemporal optimizing
approach. It states that the current account should equal the present discounted value of expected
changes in national cash flow. As such, it embodies the familiar dictum that a country should
adjust to permanent shocks but finance temporary shocks.32
To empirically implement (10), we estimate a vector autoregression (VAR) in the current
account and national cash flow for each country individually:
 (Qt  I t  Gt )  11 12   (Qt 1  I t 1  Gt 1 )    t1 

     
   2
CAt
CAt 1

  21 22  
  t 
(11)
or xt  xt 1   t . Since Et xt  k   k xt , the expression for the optimal intertemporal
consumption-smoothing current account—our proxy for the country’s external financing
need—becomes:

CAt*   (1  r )  j [1 0] k xt  [1 0]( / (1  r ))[ I   / (1  r )]1 xt
(12)
j 1
where for each country, the discount rate is set at r  0.02 , and a first-order VAR is
estimated.
32
For instance, if the shock is permanent, then by definition it is not expected to be reversed, so Δ(Qt+j - It+j Gt+j) = 0  j and, according to (10), the country should not run a current account deficit. If however there is a
purely temporary fall in output such that Δ(Qt+j - It+j - Gt+j)>0, then the country should run a deficit.
39
APPENDIX B: DATA AND SUMMARY STATISTICS
Figure B1. Colombia: Net Capital Flows to GDP (in percent), 1980-2009
(a) Threshold approach
6
(b) Cluster approach
6
Surge
4
Surge
4
2
2
Normal
0
Normal
0
Country-specific top 30th percentile
Sample top 30th percentile
Country-specific bottom 30th percentile
Sample bottom 30th percentile
-2
-4
1980
1985
1990
1995
2000
Outflow
Outflow
-2
Cut-off for surge
Cut-off for crash
-4
2005 2009
1980
1985
1990
1995
2000
2005 2009
Source: Authors’ calculations based on IFS database.
Figure B2. Surges by Region, 1980-2009
(i) Asia
80
(ii) Europe
20
16
60
80
20
16
60
12
12
40
40
8
20
4
0
0
1980 1984 1988 1992 1996 2000 2004 2008
8
20
4
0
0
1980 1984 1988 1992 1996 2000 2004 2008
Avg. net financial inflow (in % of GDP) in surge--right axis
Surge observations (in % of total regional observations)
Avg. net financial inflow (in % of GDP) in surge--right axis
Surge observations (in % of total regional observations)
(iii) Latin America and Caribbean
(iv) Middle East and Africa
20
60
60
50
40
16
40
40
30
12
8
20
20
20
10
4
0
0
1980 1984 1988 1992 1996 2000 2004 2008
Avg. net financial inflow (in % of GDP) in surge--right axis
Surge observations (in % of total regional observations)
Source: Authors’ calculations based on IFS database.
0
0
1980 1984 1988 1992 1996 2000 2004 2008
Avg. net financial inflow (in % of GDP) in surge--right axis
Surge observations (in % of total regional observations)
40
Table B1. Variable Definitions and Data Sources
Variables
Description
Source
Capital account openness
Commodity price index
Index (high=liberalized; low =closed)
Log difference betw een actual and trend commodity
price index (trend obtained from HP filter)
Index
Measures excess return per unit of risk. The index is
constructed by regressing excess returns of 64
emerging market assets on their standard deviation.
The derived coefficient from the regression gives
rise to the index w ith low er (higher) values
indicating periods of financial strain (ease)
Obtained from the intertemporal optimizing model of
the current account
First year of sovereign debt crisis (inability to pay
the principal or interest payments on the due date or
w ithin the grace period).
Chinn-Ito (2008) 1
IMF's WEO database
De facto (1=Fixed; 2=Intermediate; 3=Flexible)
Number of lenders of bank credit (0 to 15)
In billions of USD (or LC)
Net financial flow s excluding financing items and
other investment liabilities of general government (In
billions of USD)
Index
Average of 12 ICRG's political risk components
In percent
In billions of LC
Index
In USD
[(1+nominal interest rate)/(1+inflation)]-1
Difference betw een domestic real interest rate, real
US interest rate, and REER overvaluation
Log difference betw een REER and REER trend
(obtained from HP filter)
Share of countries in the region w ith a surge
IMF's AREAER
Minoiu and Reyes (2011) 4
IMF's WEO database
IMF's IFS database
Consumer price index, year average
Credit Suisse Risk Appetite Index
Optimal current account balance/GDP
Default onset
Exchange rate regime
Financial interconnectedness
GDP, current/constant prices
Net capital flow s
Nominal Effective Exchange Rate
Institutional quality index
Money market rate
Private sector credit
Real Effective Exchange Rate (REER)
Real GDP per capita
Real interest rate
Real interest rate differential
REER overvaluation
Regional contagion
Regional contagion (magnitude)
S&P 500 index
S&P 500 index volatility
Stock of foreign exchange reserves
Stock market capitalization
Trade links
U.S. 3-month Treasury Bill rate
VXO index
Net capital flow to GDP in countries in the region
experiencing a surge
Index
Annual average of tw elve-month rolling standard
deviation of S&P 500 index annual returns
In billions of USD
Value of listed shares to GDP
Calculated
as in Forbes and Warnock (2011); trade
n
links=  EX x ,i ,t 1 / GDPx ,t 1 ) * surge i ,tw here EXi,t-1 is
exportst 1from country x to i in year t-1, GDPx,t-1 is the
GDP of country x in t-1, and surgei,t is a binary
variable (=1) if country i had a surge in t
In percent
Chicago Board Options Exchange Market Volatility
Index (high values indicate greater volatility of S&P
500 index options)
IMF's WEO database
Bloomberg
Ghosh (1995) 2
Reinhart and Reinhart (2008) 3
IMF's INS database
http://w w w .prsgroup.com/Default.aspx
IMF's IFS database
IMF's IFS database
INS database
IMF's WEO database
Authors' calculations
Authors' calculations
Authors' calculations
Authors' calculations
Bloomberg
Authors' calculations
IMF's WEO database
Beck and Demirgüç-Kunt (2009) 5
Authors' calculations using bilateral
trade data from IMF's DOTS
IMF's WEO and Bloomberg
Bloomberg
1/ Chinn, M ., and H. Ito , 2008, "A New M easure o f Financial Openness," Jo urnal o f Co mparative P o licy A nalysis , Vo l. 10(3): 309-322.
2/ Gho sh, A ., 1995, "Internatio nal Capital M o bility A mo ngst the M ajo r industrial Co untries: To o Little o r To o M uch?" Eco no mic Jo urnal , Vo l. 105(428): 107-128.
3/ Reinhart, C., and Reinhart, V., 2008, "Capital Flo w B o nanzas: A n Enco mpassing View o f the P ast and P resent," NB ER Wo rking P aper 14321.
4/ M ino iu, C., and J. Reyes, 2011, “ A Netwo rk A nalysis o f Glo bal B anking: 1978-2009,” IM F Wo rking P aper WP /11/74 (Washingto n DC: IM F).
5/ B eck, T., and A . Demirgüç-Kunt, 2009, "Financial Institutio ns and M arkets A cro ss Co untries and o ver Time: Data and A nalysis," Wo rld B ank P o licy Research
Wo rking P aper No . 4943 (Washingto n DC: Wo rld B ank).
41
Table B2. List of Surge Episodes with the Threshold Approach, 1980-2009
Country
Duration a
Albania
1988-89
Albania
1997
Albania
2006-09
Argentina
1993-94
Argentina
1997-99
Armenia
1996-2000
Azerbaijan
1996-98
Azerbaijan
2003-04
Belarus
1997
Belarus
2002
Belarus
2004
Belarus
2006-07
Belarus
2009
Bosnia & Herzegovina 2001
Bosnia & Herzegovina 2003-05
Bosnia & Herzegovina 2007-08
Brazil
1980-81
Brazil
1994
Brazil
2007
Bulgaria
1993
Bulgaria
2002-08
Chile
1980-81
Chile
1990
Chile
1992-97
Chile
2008
China
1994
China
2004
Colombia
1996-97
Colombia
2007
Costa Rica
1980
Costa Rica
1999
Costa Rica
2002
Costa Rica
2006-08
Croatia
1996-97
Croatia
1999
Croatia
2001
Croatia
2003
Croatia
2006
Croatia
2008
Czech Rep.
1995
Czech Rep.
2001-02
Dominican Rep.
2000-01
Dominican Rep.
2005
Dominican Rep.
2007-08
Ecuador
1990-92
Ecuador
1994
Ecuador
1998
Ecuador
2002
Egypt
1997
Egypt
2005
Avg. net
capital flow
(% of GDP) b
9.9
5.1
7.9
8.6
5.4
12.9
27.0
32.3
4.8
5.1
4.5
6.5
8.0
15.5
13.8
12.5
6.2
7.3
6.5
18.0
26.5
13.3
8.4
7.5
5.1
4.9
5.7
5.9
4.8
6.5
4.8
5.4
8.8
12.8
14.0
11.9
11.7
12.7
11.5
11.9
10.9
6.3
4.6
5.9
9.8
5.7
6.4
5.4
4.6
8.3
Country
Duration a
Avg. net
capital flow
(% of GDP)b
El Salvador
1998
7.3
El Salvador
2003
6.7
El Salvador
2006
5.9
El Salvador
2008
7.1
Estonia
1997
16.0
Estonia
2003-04
12.9
Estonia
2006-07
16.8
Guatemala
1991-93
7.8
Guatemala
1998
5.1
Guatemala
2000-03
6.7
Hungary
1993-95
14.9
Hungary
1998-2000
11.1
Hungary
2003-06
10.6
Hungary
2008
8.8
India
2007
8.0
Indonesia
1995-96
4.6
Israel
1982
7.6
Israel
1997
6.8
Israel
1999
4.9
Israel
2008
7.5
Jamaica
1984
8.5
Jamaica
2001-02
9.8
Jamaica
2005-09
12.6
Jordan
1991-92
31.4
Jordan
2005-09
18.6
Kazakhstan
1997
11.7
Kazakhstan
2001
11.6
Kazakhstan
2004
11.0
Kazakhstan
2006
20.0
Korea
1980-81
6.0
Latvia
2004-07
21.3
Lebanon
2008-09
47.8
Lithuania
1998
11.7
Lithuania
2006-07
16.0
Macedonia
2001
11.2
Macedonia
2005
9.3
Macedonia
2007-08
11.9
Malaysia
1980-85
8.4
Malaysia
1991-93
14.3
Malaysia
1995-96
9.3
Mauritius
1988
6.1
Mauritius
1990
5.6
Mauritius
2007-09
7.1
Mexico
1980-81
6.6
Mexico
1991-93
7.7
Mexico
1997
5.2
Morocco
1994
5.5
Pakistan
2006-07
5.7
Panama
1981
7.6
Panama
1996-99
10.3
Source: Authors' calculations.
a
Refers to the years of the surge episode.
b
Mean of net capital flow to GDP (in percent) received over the surge episode.
Country
Duration a
Avg. net
capital flow
(% of GDP)b
Panama
2001
11.0
Panama
2005
13.5
Panama
2007-08
11.4
Paraguay
1980-82
7.1
Paraguay
2005
5.2
Paraguay
2007
6.4
Peru
1994-97
7.9
Peru
2002
4.9
Peru
2007-08
10.5
Philippines
1980
6.8
Philippines
1991
5.6
Philippines
1994-97
9.9
Philippines 1999-2000
4.8
Poland
1980
4.9
Poland
1995-2000
6.7
Poland
2003
4.6
Poland
2005
7.1
Poland
2007-09
8.3
Romania
2001-02
7.2
Romania
2004-08
14.3
Russia
2007
7.6
Serbia
2007
19.1
Slovak Rep.
1996
11.3
Slovak Rep.
2002
22.7
Slovak Rep. 2004-05
13.2
South Africa
1997
5.4
South Africa 2005-07
6.0
South Africa
2009
5.8
Sri Lanka
1980
5.0
Sri Lanka
1982
6.5
Sri Lanka
1989
5.0
Sri Lanka
1993-94
6.5
Thailand
1981
6.2
Thailand
1988-96
10.0
Tunisia
1981-82
6.1
Tunisia
1984
5.4
Tunisia
1992-93
5.6
Tunisia
2006
9.5
Tunisia
2008-09
6.5
Turkey
1993
4.5
Turkey
2004-08
6.9
Ukraine
2005
9.4
Ukraine
2007-08
7.8
Uruguay
1980-83
6.2
Uruguay
2005-08
9.3
Venezuela
1990
29.2
Venezuela
1992-93
4.7
Vietnam
1996
11.8
Vietnam
2007-09
17.1
42
Table B3. List of Surge Episodes with the Cluster Approach, 1980-2009
Durationa Average net
capital flow s
(% of GDP)b
Albania
1988-89
9.87
Albania
1997
5.14
Albania
2006-09
7.95
Algeria
1980
2.28
Algeria
1989
1.53
Algeria
2008-09
3.79
Argentina
1992-94
6.91
Argentina
1996-99
5.15
Armenia
1996-2000
12.88
Azerbaijan
1996-98
26.96
Azerbaijan
2003-04
32.29
Belarus
2002
5.10
Belarus
2007
8.35
Belarus
2009
7.99
Bosnia & Herzegovina
2001
15.53
Bosnia & Herzegovina 2004-05
14.55
Bosnia & Herzegovina
2007
13.63
Brazil
1980-82
5.29
Brazil
1994-96
5.03
Brazil
2000-01
4.01
Brazil
2007
6.45
Brazil
2009
4.43
Bulgaria
1993
17.95
Bulgaria
2002
22.80
Bulgaria
2005-08
33.74
Chile
1980-81
13.34
Chile
1990
8.40
Chile
1992-94
7.67
Chile
1996-97
8.40
China
1993-96
4.26
China
2003-04
4.53
China
2009
2.85
Colombia
1982
3.55
Colombia
1985
4.03
Colombia
1993-97
4.75
Colombia
2007-08
4.16
Costa Rica
1980
6.49
Costa Rica
1995
4.31
Costa Rica
1999
4.82
Costa Rica
2002-03
4.66
Costa Rica
2005-08
7.73
Croatia
1997
14.16
Croatia
1999
13.97
Croatia
2001
11.85
Croatia
2006
12.70
Czech Rep.
1995
11.89
Czech Rep.
2002
14.17
Dominican Rep.
1980
4.44
Dominican Rep.
1998-01
5.16
Dominican Rep.
2005
4.60
Dominican Rep.
2007-09
5.16
Ecuador
1990-92
9.80
Ecuador
1994
5.67
Ecuador
1998-98
6.37
Ecuador
2002
5.36
Egypt
1997-98
3.96
Country
Durationa Average net
capital flow s
(% of GDP)b
Egypt
8.34
2005
Egypt
4.08
2007-08
El Salvador
4.13
1995
El Salvador 1997-99
5.31
El Salvador 2002-03
5.02
El Salvador 2005-06
5.10
El Salvador 2008-08
7.14
Estonia
15.96
1997
Estonia
13.16
2004
Estonia
16.84
2006-07
Guatemala 1991-93
7.84
Guatemala 1998-98
5.11
Guatemala 2000-03
6.70
Hungary
14.94
1993-95
Hungary
11.11
1998-2000
Hungary
11.51
2004-06
Hungary
8.83
2008-08
India
3.24
1994
India
3.43
2003-04
India
6.02
2006-07
India
3.30
2009
Indonesia
3.17
1990-93
Indonesia
4.59
1995-96
Israel
5.98
1981-82
Israel
4.33
1987
Israel
5.08
1995-97
Israel
4.49
1999-2000
Israel
7.51
2008-08
Jamaica
8.53
1984
Jamaica
9.82
2001-02
Jamaica
12.58
2005-09
Jordan
31.38
1991-92
Jordan
18.57
2005-09
Kazakhstan
11.68
1997
Kazakhstan
11.58
2001
Kazakhstan
11.05
2004
Kazakhstan
19.96
2006
Korea, Rep. 1980-82
5.14
Korea, Rep. 1995-96
3.82
Korea, Rep. 2003-03
3.19
Korea, Rep.
4.15
2009
Latvia
23.60
2005-07
Lebanon
47.85
2008-09
Lithuania
11.68
1998-98
Lithuania
15.97
2006-07
Macedonia, F 2001
11.23
Macedonia, F 2005
9.34
Macedonia, F 2007-08
11.94
Malaysia
9.06
1981-85
Malaysia
14.32
1991-93
Malaysia
9.28
1995-96
Mauritius
3.74
1980
Mauritius
6.12
1988-88
Mauritius
5.60
1990
Mauritius
4.26
1999
Mauritius
7.15
2007-09
Country
Source: Authors' calculations.
a
Refers to the years of the surge episode.
b
Mean of net capital flow s to GDP (in percent) received over the surge episode.
Durationa Average net
capital flow s
(% of GDP) b
Mexico
1980-81
6.62
Mexico
1991-93
7.68
Mexico
1997
5.18
Morocco
1990-94
3.74
Morocco
1999
3.22
Morocco
2004
2.50
Pakistan
1993-94
3.10
Pakistan
1996
3.62
Pakistan
2005-08
4.45
Panama
1981
7.55
Panama
1996-99
10.32
Panama
2001
10.99
Panama
2005
13.55
Panama
2007-08
11.37
Paraguay
1980-82
7.14
Paraguay
2005
5.17
Paraguay
2007-09
4.88
Peru
1994-97
7.94
Peru
2007-08
10.51
Philippines
1980
6.80
Philippines
1991
5.63
Philippines
1994-97
9.91
Philippines
1999
5.17
Poland
1995-96
7.39
Poland
1998-2000
6.81
Poland
2005
7.13
Poland
2007-09
8.34
Romania
2002
7.86
Romania
2004-08
14.26
Russian Fed. 1997
3.43
Russian Fed. 2002
4.13
Russian Fed. 2005-07
4.42
Serbia
2007
19.11
Slovak Rep.
2002
22.67
Slovak Rep.
2005
15.03
South Africa 1997-98
4.23
South Africa 2004-09
5.24
Sri Lanka
1980
5.03
Sri Lanka
1982
6.47
Sri Lanka
1989
5.02
Sri Lanka
1993-94
6.49
Thailand
1989-96
10.51
Tunisia
1981-82
6.13
Tunisia
1984
5.43
Tunisia
1993
6.47
Tunisia
2006
9.50
Tunisia
2008-09
6.54
Turkey
1993
4.51
Turkey
2004-08
6.86
Ukraine
2005
9.43
Ukraine
2007-08
7.80
Uruguay
1982
10.37
Uruguay
2006-08
10.70
Venezuela
1987
3.45
Venezuela 1990-93
10.40
Vietnam
2007-09
17.11
Country